AI Coding Tools Double Outages at Fragile Software Firms, Tacho Finds

Data vs Hype: How Orgs Actually Win with AI - The Pragmatic Summit — The Pragmatic Engineer

The Pragmatic Engineer

Industry benchmark data reveals generative AI acts as an organizational multiplier rather than an equalizer, halving customer incidents in disciplined engineering teams while doubling outages across fragile codebases.

AI-assisted summary of Data vs Hype: How Orgs Actually Win with AI - The Pragmatic Summit — original source. Check the original for full context.

Source published:

Key ideas

  • Industry telemetry across 121,000 developers shows 92.6% monthly AI tool adoption and an average time savings of 4.08 hours per week, with upstream merged code written by AI rising to 26.9% by February 2026. (3:24)
  • AI adoption has cut developer time to their 10th pull request by 50% across two years, a ramp-up velocity that Microsoft research shows persists for a developer's first two years of organizational tenure. (6:06)
  • AI functions as an organizational amplifier rather than an equalizer: mature teams halved customer-facing incidents, whereas low-maturity teams with fragile processes saw outages double. (7:45)
  • A July 2025 MIT study revealed high AI tool adoption alongside low business transformation, demonstrating that task-level developer automation fails to move corporate profit and loss statements without systemic change management. (9:54)
  • Autonomous agent usage is expanding rapidly, with OpenAI reporting that internal developers using Codex daily ship roughly 60% more pull requests per week than peers using conventional AI tools. (15:22)

Detailed notes

## Background Published on YouTube by The Pragmatic Engineer on February 24, 2026, this keynote from The Pragmatic Summit features CTO Laura Tacho examining how engineering organizations can move past artificial intelligence marketing hype to achieve genuine business value. The session accompanied an in-depth analysis on the future of software engineering published in The Pragmatic Engineer newsletter.

## TL;DR Drawing on benchmarks from over 121,000 developers across 450 companies, Laura Tacho shows that while AI coding assistant adoption has reached 92.6% and saves engineers roughly four hours weekly, aggregate averages hide a stark organizational divide. High-maturity teams with strong continuous integration and testing cut production incidents by 50%, whereas fragile codebases merely automate technical debt and double their outages. Tacho argues that treating AI merely as an individual desk tool limits productivity, urging leaders to reject indiscriminate 'spray and pray' licensing and instead apply agentic workflows directly to systemic developer experience bottlenecks and real customer pain points.

## Atmosphere & style The presentation strikes a refreshing balance between visionary enthusiasm and sharp, data-backed pragmatism. Tacho commands the stage with an engaging, direct narrative style, weaving mid-century Apollo space race analogies and wry humor about unhinged open-source experiments alongside hard telemetry and research from MIT and Microsoft. It feels like an executive reality check that honors the genuine joy developers feel while programming with agents, yet soberly challenges leaders to do the unglamorous architectural work required to make those tools succeed.

## Bulletin - [03:24](https://www.youtube.com/watch?v=LOHgRw43fFk&t=204s) Industry telemetry across 121,000 developers shows 92.6% monthly AI tool adoption and an average time savings of 4.08 hours per week, with upstream merged code written by AI rising to 26.9% by February 2026. - [06:06](https://www.youtube.com/watch?v=LOHgRw43fFk&t=366s) AI adoption has cut developer time to their 10th pull request by 50% across two years, a ramp-up velocity that Microsoft research shows persists for a developer's first two years of organizational tenure. - [07:45](https://www.youtube.com/watch?v=LOHgRw43fFk&t=465s) AI functions as an organizational amplifier rather than an equalizer: mature teams halved customer-facing incidents, whereas low-maturity teams with fragile processes saw outages double. - [09:54](https://www.youtube.com/watch?v=LOHgRw43fFk&t=594s) A July 2025 MIT study revealed high AI tool adoption alongside low business transformation, demonstrating that task-level developer automation fails to move corporate profit and loss statements without systemic change management. - [15:22](https://www.youtube.com/watch?v=LOHgRw43fFk&t=922s) Autonomous agent usage is expanding rapidly, with OpenAI reporting that internal developers using Codex daily ship roughly 60% more pull requests per week than peers using conventional AI tools.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=LOHgRw43fFk&t=4s) The current AI era mirrors the mid-twentieth-century space race, balancing immense excitement and technical wonder against valid skepticism regarding cost and return on investment. - [01:14](https://www.youtube.com/watch?v=LOHgRw43fFk&t=74s) Vendor claims of 100% developer productivity surges conflict with mixed empirical realities, making data-backed measurement essential to cut through marketing hype. - [03:24](https://www.youtube.com/watch?v=LOHgRw43fFk&t=204s) Adoption has largely plateaued at near-ubiquitous levels, with approximately 75% of surveyed engineers using AI coding assistants weekly. - [03:24](https://www.youtube.com/watch?v=LOHgRw43fFk&t=204s) The proportion of upstream production code authored directly by AI without substantial human intervention surpassed 30% among daily active users. - [06:06](https://www.youtube.com/watch?v=LOHgRw43fFk&t=366s) Onboarding speed has doubled industry-wide since 2024, lowering cognitive friction for junior developers, internal lateral transfers, and non-technical contributors. - [07:45](https://www.youtube.com/watch?v=LOHgRw43fFk&t=465s) Statistical industry averages mislead executives because AI radically polarizes engineering outcomes based on preexisting system hygiene and architectural maturity. - [09:54](https://www.youtube.com/watch?v=LOHgRw43fFk&t=594s) The GenAI divide stems from treating AI as an individual desk tool rather than addressing end-to-end organizational change management. - [12:10](https://www.youtube.com/watch?v=LOHgRw43fFk&t=730s) Like space exploration spinning off pragmatic inventions for Earth, high-risk agentic experimentation must ultimately be translated into grounded solutions for customer problems. - [16:55](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1015s) Real-world agent implementations demonstrate strong operational gains, such as Haven tripling healthcare CSAT and Cisco halving code review times across 18,000 engineers. - [19:25](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1165s) Achieving consensus among specialized, interacting agents is emerging as one of the primary architectural and algorithmic engineering challenges of 2026. - [20:14](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1214s) Artificial intelligence cannot fix underlying systemic or human dysfunction; migrating flawed engineering processes to agentic workflows simply accelerates the accumulation of technical debt. - [22:12](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1332s) Indiscriminate 'spray and pray' tool licensing fails; organizations must adopt structured measurement frameworks auditing adoption, organizational impact, and financial cost. - [24:04](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1444s) Foundational developer experience practices—like clean interfaces, thorough documentation, and fast CI pipelines—are the mandatory operational prerequisites for agent experience. - [26:39](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1599s) Primary blockers to enterprise AI readiness are organizational rather than technical, driven by weak executive sponsorship, poor developer UX, and lacking change management. - [29:03](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1743s) Engineering leaders must maintain a balanced approach that combines long-term technological vision with disciplined, pragmatic execution focused on current business realities.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=LOHgRw43fFk&t=4s) The Space Race Analogy for the AI Era

Laura Tacho opens her keynote by emphasizing the need for a grounded, pragmatic discussion about what is genuinely taking place with artificial intelligence inside engineering organizations, what leaders should anticipate, and how autonomous agents are reshaping software development. To anchor this down-to-earth perspective, she turns to an unexpected parallel: the mid-twentieth-century space race and the age of exploration. Tacho recounts a conversation she had the previous week with both a CTO and co-founder of a nimble startup and a principal engineering lead at a massive, highly regulated bank. Despite operating at opposite ends of the corporate spectrum, both leaders spent fifteen minutes marveling at what their teams were building, sharing a palpable sense that AI had restored the sheer joy of coding.

With ideas multiplying faster than teams can implement them, Tacho notes that the current technical climate mirrors Carl Sagan’s famous observation: somewhere, something incredible is waiting to be known. Just as the Apollo era unlocked an intoxicating sense of possibility about reaching the Moon and Mars, today’s developer community is swept up in the thrill of discovery and technical experimentation. Yet, as Tacho points out, historical space exploration was never met with universal acclaim; it sparked deep, persistent skepticism from critics who questioned the astronomical financial investments when pressing socioeconomic crises remained unsolved on Earth.

- Tacho observed identical enthusiasm across both early-stage startups and heavily regulated tier-one financial institutions, where leaders reported that AI tooling had actively brought back the fun and joy of software development. - The parallel to the space age highlights that transformative technological eras are rarely purely scientific; they operate at the intersection of global politics, economics, and human ambition.

> “There is so much to learn and so much to build.” — Laura Tacho

---

[01:14](https://www.youtube.com/watch?v=LOHgRw43fFk&t=74s) Wonder vs. Skepticism: Balancing Exploration with Reality

Exploring the universe was not a silver bullet for human suffering, but the Apollo Moon landing nonetheless represented an undeniable, defining inflection point that permanently expanded humanity's collective imagination. Tacho argues that the technology sector finds itself in an identical tension regarding AI. On one hand, the industry is flooded with unbridled optimism and marketing claims: assertions of 100% productivity surges, promises that software will soon be written entirely by synthetic agents, and speculation about the imminent emergence of the world's first single-person, billion-dollar startup powered by an individual and an army of autonomous AI workers.

On the other hand, corporate decision-makers and empirical researchers are voicing legitimate skepticism. Organizations are grappling with the immense financial expense of running large models, the mounting environmental toll of data centers, and contradictory research studies evaluating whether AI genuinely makes developers faster. In certain constrained workflows, AI serves as a powerful accelerator; in others, it introduces friction, technical debt, and context-switching that actually degrades overall velocity. Tacho stresses that engineering leaders do not need to extinguish their sense of wonder, but they must balance creative curiosity with strict pragmatism, resisting unverified vendor hype by turning directly to empirical measurement.

> “We need to keep the sense of wonder while also balancing it with pragmatism and beating the hype by looking at data.” — Laura Tacho

---

[03:24](https://www.youtube.com/watch?v=LOHgRw43fFk&t=204s) New Industry Benchmarks in 2026

To replace speculation with hard evidence, Tacho unveils a set of freshly gathered industry benchmarks collected between November 1, 2025, and February 1, 2026. Drawing from a telemetry and survey pool of 121,000 developers across more than 450 companies, the data illustrates that AI tool adoption has largely plateaued at near-ubiquitous levels. Approximately 92.6% of professional engineers report leveraging an AI coding assistant—predominantly tools like Cursor, GitHub Copilot, Anthropic’s Claude, and OpenAI Codex, rather than generic web-based ChatGPT—at least once a month, while roughly 75% integrate these tools into their weekly routine.

```mermaid flowchart TD accTitle: 2026 AI Developer Benchmarks accDescr: Adoption rates, weekly time savings, and production code authorship percentages across engineering organizations.

subgraph Adoption ["Adoption (121k Developers)"] A1["92.6% Monthly Usage"] A2["75.0% Weekly Usage"] end

subgraph Velocity ["Time Savings"] B1["4.08 Hours Saved / Week"] B2["~10% Total Developer Time"] end

subgraph Production ["Code Merged (42.6k Developers)"] C1["26.9% AI-Authored Code"] C2[">30% for Daily AI Users"] end

Adoption --> Velocity Velocity --> Production ```

When evaluating developer velocity, self-reported time savings sit at an average of 4.08 hours saved per developer per week. Tacho highlights that while time savings is merely an early leading indicator rather than a comprehensive metric for engineering productivity, the figure has remained steady, hovering around the four-hour mark across multiple quarters (following 3.6 to 3.7 hours in Q4 2025). This aligns closely with public disclosures from companies like Google citing an approximate 10% net productivity lift. Where the metric is accelerating dramatically, however, is in AI-authored code: measuring code merged into upstream, customer-facing branches without substantial human intervention. Across a targeted sample of 42,600 developers, AI-authored production code rose from 22% in late 2025 to 26.9% by February 2026, with daily active AI users now seeing more than 30% of their shipped code produced directly by assistants.

- **Adoption Base**: 121,000 developers across 450+ organizations surveyed between November 2025 and February 1, 2026. - **Usage Frequency**: 92.6% active at least monthly; ~75% active at least weekly. - **Reported Time Saved**: 4.08 hours per engineer per week, representing roughly a 10% gain, consistent with historical quarterly data since Q2 2025. - **Production Code Footprint**: 26.9% of all upstream merged code is AI-authored without major manual intervention, up sharply from 22% in Q4 2025. - **Power User Threshold**: Developers who use AI tools daily now see over 30% of their merged pull requests authored by synthetic assistants.

> “What is changing and what is moving up very quickly is the amount of code getting merged upstream or in a customer-facing environment that was written by AI that was merged without significant human intervention.” — Laura Tacho

---

[06:06](https://www.youtube.com/watch?v=LOHgRw43fFk&t=366s) Halving Onboarding Speed: The Time to 10th PR

Tacho identifies employee onboarding as one of the most compelling and substantiated operational applications for AI assistants. Having long suspected that conversational coding agents could dramatically streamline how newcomers navigate complex documentation and foreign architectures, she tracked cross-industry onboarding velocity from Q1 2024 through Q4 2025. The metric monitored was "time to 10th pull request"—a widely accepted milestone in developer experience research signaling that an engineer has moved past superficial setup tasks and achieved baseline operational autonomy. Over that two-year span, the time required for a developer to ship their tenth pull request was cut exactly in half, tracking in near-perfect lockstep with the industry-wide adoption curve of AI tooling.

Critically, this acceleration is not limited to junior or newly hired staff; the same dramatic ramp-up occurs when veteran engineers pivot between disparate internal services or when non-technical collaborators begin contributing to repositories. Emphasizing the compounding value of this shift, Tacho cites empirical research conducted by Brian Houck at Microsoft, a co-author of the SPACE framework for developer productivity. Houck’s investigation demonstrated that the velocity an engineer achieves while working toward their tenth pull request establishes a persistent performance trajectory that endures for their first two years of employment at the company.

- **Onboarding Velocity Reduction**: Time to reach the 10th merged pull request dropped by 50% between Q1 2024 and Q4 2025. - **Long-Term Performance Impact**: Microsoft research by Brian Houck shows that onboarding velocity gains stick with developers for at least twenty-four months of their organizational tenure. - **Cognitive Load Relief**: The primary mechanism driving faster onboarding is AI's ability to contextualize codebases, surface institutional knowledge rapidly, and reduce discovery friction.

> “So if you onboard faster that productivity gain isn't just onboarding it actually sticks with them for at least two years after they have started at the company.” — Laura Tacho

---

[07:45](https://www.youtube.com/watch?v=LOHgRw43fFk&t=465s) Why Averages Mislead: Divergent Quality and the Big Bang Effect

Despite sharing these aggregate benchmarks, Tacho issues an urgent warning to engineering executives: industry averages are purely mathematical artifacts that often obscure reality. When underlying data points polarize and diverge toward extreme opposite ends of a spectrum, the statistical mean remains deceptively stationary. Consequently, an average does not represent a standard, typical, or predictable outcome for any specific engineering organization. In reality, there is no universal or uniform experience with developer AI, because the technology's impact is heavily dictated by existing internal workflows, tooling maturity, and company culture.

To illustrate this divergence, Tacho invokes the Big Bang. In the early moments of the cosmos, an explosive release of centralized energy propelled matter outwards, causing the physical distance between celestial bodies to rapidly widen. The introduction of AI into enterprise software development functions as an identical explosive catalyst. Because organizational performance is multi-dimensional, AI acts not as an equalizer, but as an amplifier and multiplier that magnifies preexisting institutional strengths and weaknesses, thrusting companies into radically different performance tiers.

```mermaid flowchart LR accTitle: AI as an Organizational Amplifier accDescr: How introducing AI multiplies pre-existing engineering hygiene into starkly divergent outcomes.

A["AI Tool Adoption"] --> B{"Baseline System Hygiene"}

subgraph HighMaturity ["Mature Engineering Culture"] B -->|"Strong CI/CD & Rigor"| C["50% Incident Reduction"] C --> C1["Higher Code Maintainability"] C --> C2["Higher Change Confidence"] end

subgraph LowMaturity ["Fragile Systems & Debt"] B -->|"Weak Reviews & Testing"| D["2x Customer-Facing Incidents"] D --> D1["Compounded Technical Debt"] D --> D2["Amplified Instability"] end ```

Tacho substantiates this polarity using incident data gathered from over 67,000 developers between November 2025 and February 2026. While the baseline metrics might appear stable on paper, production software quality is splitting into stark extremes across organizations:

| Engineering Quality Metric | High-Maturity Organizations | Fragile / Low-Maturity Organizations | |:--- |:--- |:--- | | **Customer Incident Volume** | **50% reduction** in production outages | **200% increase (2x)** in customer incidents | | **System Health Response** | Amplified testing rigor, high change confidence | Faster deployment of unvetted regressions | | **Maintainability Trajectory** | Improved code clarity and automated validation | Compounded technical debt and review fatigue |

Organizations equipped with disciplined testing frameworks, continuous integration, and mature peer-review cultures have harnessed AI to accelerate while simultaneously halving their customer-impacting outages. Conversely, organizations plagued by fragile architectures and unstructured deployment practices have merely automated the generation of technical debt, doubling their incident rates by shipping lower-quality code at higher frequencies.

> “One thing that is absolutely true, one thing that is common is that there is no typical experience with AI.” — Laura Tacho

[09:54](https://www.youtube.com/watch?v=LOHgRw43fFk&t=594s) High Adoption, Low Transformation and the GenAI Divide

Organizations that were already struggling before the advent of generative AI are discovering that new tooling has not resolved their underlying dysfunction; instead, it has simply enabled them to be dysfunctional faster. Across the industry, the economic returns of AI remain starkly uneven, characterized by steep drop-offs as initiatives attempt to transition from pilot experiments into production and ultimately connect to profitability. Laura Tacho references an MIT study published in July 2025 titled *The GenAI Divide*, which surveyed 152 organizations and revealed a fundamental disconnect: the industry is experiencing exceptionally high tool adoption—mirroring figures like GitHub's 92.6% and DORA research hovering around 90%—alongside remarkably low organizational transformation.

```mermaid flowchart LR accTitle: The GenAI Adoption vs Impact Disconnect accDescr: Diagram illustrating how applying AI solely to individual desk tasks yields low productivity, whereas organizational transformation drives business profit.

subgraph Narrow["Task-Level Surface Area"] A["Individual Developer at Desk"] --> B["Automated Coding Tasks"] B --> C["Low Productivity Ceiling"] end

subgraph Org["Systems-Level Change"] D["Org Change Management"] --> E["Cross-Functional Workflows"] E --> F["P&L & Bottom-Line Profit"] end

Narrow -.->|"Fails to bridge to"| Org ```

The core obstacle is that real transformation is deeply uncomfortable. Companies that previously abandoned their cloud or agile transformations are now quietly giving up on their AI transformations because overhauling entrenched organizational habits requires painful, systemic change management. The prevailing hype promised that developers could simply experiment at their desks and magically generate bottom-line profit, but confining AI strictly to individual coding tasks produces a remarkably low productivity ceiling. As Laura Tacho argues, “If we want organizational results we have to think about it on an organizational level not on a coding task level.”

- **Survey Scope:** The July 2025 MIT study, *The GenAI Divide*, evaluated 152 organizations navigating the gap between AI pilots and production. - **The Adoption Gap:** Developer adoption rates consistently hover around 90% to 92.6%, yet most enterprises fail to see a corresponding movement in their profit and loss statements. - **The Core Fallacy:** Treating AI as an individual developer tool rather than an organizational change management challenge caps potential productivity gains.

---

[12:10](https://www.youtube.com/watch?v=LOHgRw43fFk&t=730s) The Expanding Universe: Fun, Moonshots, and Earthbound Value

The scope of what software teams can accomplish is broadening through agentic workflows, which bring an explosion of experimentation, hype, and genuine possibility. Laura Tacho draws a direct parallel to the 1960s Apollo space race: while cultural imagination promised flying cars and civilization living on the moon, the real triumph of space exploration was developing systems-level innovations that transformed everyday life on Earth—such as quartz watches, barcodes, sunglasses, and space blankets. In the contemporary AI ecosystem, developers are similarly building wild, boundary-pushing experiments—from open-source projects like Gas Town (which Tacho humorously warns is utterly "unhinged"), Moltbot, and Clawbot, to autonomous Ralph loops and quick personal apps spun up at a nail salon.

However, tinkering with color-matching apps does not equate to a multinational financial institution moving its revenue needle. Reflecting on discussions from an engineering retreat with Martin Fowler and Kent Beck, Tacho explains that the industry is grappling with how to reconcile exploratory innovation with corporate profit and loss statements. While not every enterprise needs to live on the technological frontier, embracing bold experimentation is what expands the boundaries of problem-solving. As Laura Tacho summarizes: “Not everyone goes to the moon, and it's okay not to go to the moon.” The true mandate for businesses is taking the lessons gleaned from that exploratory frontier and applying them directly to systemic challenges back home.

- **Wild Experimentation:** Platforms and community projects like Gas Town, OpenClaw, Moltbot, and Ralph loops demonstrate the immense creative energy of the current developer zeitgeist. - **The Space Race Analogy:** Moonshot exploration is justified not because everyone relocates to the moon, but because the byproduct technologies solve hard, systemic problems on the ground. - **Business vs. Wonder:** While curiosity fuels engineering milestones, commercial enterprises must deliberately bridge exploratory agents to balance-sheet economics.

---

[15:22](https://www.youtube.com/watch?v=LOHgRw43fFk&t=922s) Early Telemetry on Agents and the Rise of Codex

While broad industry telemetry on agentic workflows is still nascent, early data from pioneering engineering organizations reveals that autonomous agents are rapidly becoming foundational to daily work. Drawing on telemetry from roughly 3,000 developers across six forward-leaning companies, Tacho highlights that 80% of these engineers interact with agentic workflows at least once a week, with more than 50% relying on them every single day to ship code.

```mermaid flowchart TD accTitle: Codex Velocity and Adoption Telemetry accDescr: Telemetry metrics showing Codex adoption growth, weekly developer usage, and internal PR shipping velocity.

A["Codex Desktop App<br>(Launched Feb 2)"] --> B["Over 1M Downloads<br>& +60% Weekly User Growth"] B --> C["Trillions of Tokens<br>Processed Weekly"]

subgraph Internal["OpenAI Internal Developer Metrics"] D["95% of Developers<br>Use Codex Daily"] E["60% More PRs Shipped<br>per Week vs. Other AI Tools"] end

C -.-> Internal ```

This momentum is underscored by the explosive release of the Codex desktop application on February 2nd and the subsequent rollout of GPT-5.3 Codex. Within weeks of launch, the client surpassed one million downloads, experiencing a 60% surge in active users over a single week while processing trillions of tokens weekly. Inside OpenAI, where 95% of engineering teams use Codex internally to deliver features, engineers leveraging Codex ship approximately 60% more pull requests per week compared to those using conventional AI coding tools. This metric demonstrates the unusually high ceiling of possibility when tooling evolves from passive code autocompletion to active agentic execution.

- **Agent Telemetry Cohort:** In a telemetry study of 3,000 developers at six early-adopting enterprises, over half incorporate autonomous agents into their daily programming routines. - **Codex Growth Trajectory:** Following its February 2nd debut, the Codex desktop app cleared 1,000,000 downloads and logged a 60% week-over-week user expansion. - **Engineering Throughput:** OpenAI's internal developers using Codex realize a 60% lift in merged pull requests per week over peers relying on older AI utilities.

---

[16:55](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1015s) Real-World Case Studies: Haven, Cisco, and JPMorgan Chase

To illustrate how agentic workflows provide value across different corporate scales, Tacho contrasts an agile healthcare startup with large-scale enterprise deployments. In San Francisco, Haven Headache and Migraine Center set out to provide remote neurological care, utilizing agentic development to differentiate strictly between durable architectural code and disposable prototypes. Rather than generating low-quality AI slop, Haven uses Ralph loops combined with design artifacts from Linear and Figma—converted into structured PRD JSON files—to autonomously produce working patient portal prototypes complete with comprehensive test suites and documentation. Concurrently, Haven fine-tuned a HIPAA-compliant model on hundreds of thousands of inbound symptom logs to intelligently route patient messages for prescription refills and follow-ups. This initiative elevated customer satisfaction to three times the industry average and yielded tangible clinical results, reducing both the frequency and severity of patient headache days.

```mermaid flowchart LR accTitle: Haven Healthcare Agentic Prototyping Pipeline accDescr: Pipeline showing how Haven converts Linear and Figma assets into PRD JSON, executes Ralph loops, and outputs production-grade prototypes.

A["Linear & Figma<br>Artifacts"] --> B["PRD Converted<br>to JSON"] B --> C["Autonomous<br>Ralph Loops"] C --> D["High-Quality Prototypes<br>with Docs & Tests"] D --> E["Clinical Care Routing<br>& 3x Industry CSAT"] ```

At enterprise scale, companies are deploying agents to conquer legacy complexity and friction. A major industrial manufacturing enterprise leveraged GitHub Copilot and Claude to build an internal developer portal that streamlined onboarding across disparate teams. At Cisco, an organization of 18,000 engineers uses Codex daily to manage intricate codebase migrations and automate code reviews, which has cut total code review turnaround time by 50%. In financial services, JPMorgan Chase published research on their Multi-Agent Framework for Annotation (MAFA), proving that structured multi-agent coordination can conquer high-stakes data classification tasks in heavily regulated environments.

- **Haven Headache and Migraine Center:** Uses autonomous Ralph loops to rapidly prototype patient portals from Linear and Figma specs, generating well-tested code faster than manual builds. - **Clinical and Operational Impact:** Haven's HIPAA-compliant triage model, trained on hundreds of thousands of symptom logs, drove customer satisfaction to 3x the healthtech baseline while measurably lowering patient headache frequency. - **Cisco's 18,000-Seat Footprint:** Broad deployment of Codex for continuous migration work and peer review cut code review cycles by 50%. - **Enterprise Precedents:** Enterprise manufacturing organizations are building internal dev portals with Copilot and Claude, while JPMorgan Chase has open-sourced architectural approaches through its MAFA (Multi-Agent Framework for Annotation) research paper.

[19:25](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1165s) Multi-Agent Workflows and the Consensus Challenge

Laura Tacho highlights emerging architectures where organizations build out an entire ecosystem of specialized agents operating in true multi-agent workflows, drawing parallels to systems like Steve Yegge's Gas Town. In these advanced setups, individual agents are assigned dedicated functions: some intake interactions and annotate their intent (such as classifying whether an inquiry is an FAQ), while an entirely separate tier of agents reranks, calibrates, and validates the output. Because independent agents often generate conflicting evaluations or divergent outputs, introducing formal consensus algorithms becomes mandatory. As Laura Tacho emphasized:

> “I believe consensus among agents is going to be a huge problem to solve in 2026.”

```mermaid flowchart TD accTitle: Multi-Agent Consensus Architecture accDescr: Workflow showing raw interactions flowing to annotating agents, then calibrating agents, and resolving through consensus algorithms.

subgraph Ingestion ["Ingestion & Annotation"] A["User Interactions"] --> B["Annotator Agents<br/>(Intent, FAQ, Context)"] end

subgraph Validation ["Validation Tier"] B --> C["Reviewer Agents<br/>(Rerank & Calibrate)"] B --> D["Validator Agents<br/>(Check Constraints)"] end

subgraph Resolution ["Resolution"] C --> E["Consensus Algorithms"] D --> E E --> F["Validated Output"] end ```

[20:14](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1214s) AI Won’t Fix Itself: The Agile 25th Anniversary Retreat

Laura Tacho recounts attending the Future of Software Development retreat organized by Martin Fowler and Thoughtworks to mark the 25th anniversary of the Agile Manifesto. High in the mountains, she spent a day and a half discussing agents with industry leaders including Gergely Orosz, Steve Yegge, and Kent Beck, examining how agentic systems can be deployed responsibly, sustainably, and ethically. Despite intense experimentation with cutting-edge architectures, the group reached a sobering collective realization: artificial intelligence cannot magically cure structural dysfunctions. Summing up an impromptu discussion between herself, Kent Beck, and Steve Yegge, Laura Tacho argued:

> “Organizations are constrained by human and systems level problems. We remain skeptical of the promise of any technology to improve organizational performance without first addressing those human and systems level constraints.”

She cautioned that ignoring existing systemic flaws merely transports organizational dysfunction to a new frontier:

> “The risk is if we don't address the systems level problems, we will just take them to space with us.”

She compared this illusion to space colonization: moving to the moon does not eliminate trash, pollution, or traffic. Consequently, engineering leaders must shift their focus away from chasing speculative science fiction toward applying AI deliberately to solve underlying systemic bottlenecks.

[22:12](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1332s) Winning Strategy 1: Measure Goals, Avoid “Spray and Pray”

Organizations achieving tangible success with AI do not distribute developer licenses indiscriminately and hope for productivity miracles. Denouncing this widespread tactic, Laura Tacho was unequivocal:

> “Spray and prey does not work.”

Success requires targeting experimentation at clearly defined problems, establishing baseline goals, and quantitatively auditing outcomes. Quoting Spock’s adage that *"insufficient facts always invite danger,"* she noted that measuring developer productivity and engineering excellence remains notoriously complex.

To give teams a concrete measurement path, she introduced the AI Measurement Framework, which she co-authored with Abi Noda, CEO of DX. Designed to complement the Core 4 framework, it structures evaluation around three pillars:

- **Adoption & Usage**: Tracking daily utilization, prompt depth, and developer workflow engagement rather than simple seat allocation. - **Organizational Impact**: Mapping adoption directly to the Core 4 impact dimensions—speed, developer experience, code quality, and innovation ratio. - **Cost**: Monitoring escalating licensing, model compute, and token costs to evaluate whether the financial investment remains justifiable as tool pricing climbs.

[24:04](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1444s) Winning Strategy 2: Developer Experience Matters More Than Ever

Laura Tacho offered an unconventional tactic for securing engineering resources: take any developer experience initiative you need funded, pitch it to executive leadership as “agent experience,” and watch the budget get approved immediately. While humorous, the underlying reality is profound: the exact technical hygiene practices engineers have spent decades begging for—clean service boundaries, thorough documentation, comprehensive test suites, fast CI feedback loops, and robust quality gates—are the exact prerequisites that allow autonomous agents to operate effectively.

```mermaid flowchart LR accTitle: Developer Experience as Agent Experience accDescr: Legacy DevX practices directly enabling autonomous agent workflows.

subgraph Legacy ["Foundational DevX"] A["Comprehensive Documentation"] B["Fast CI & Short Feedback Loops"] C["Strict Testing & Quality Gates"] end

subgraph Agentic ["Agent Readiness"] A --> D["Accurate Agent Context"] B --> E["Rapid Multi-Step Iteration"] C --> F["Autonomous Validation"] end ```

Looking at current industry benchmarks, AI tooling hovers around saving approximately 4 hours per week. That modest time savings from faster code typing is completely erased if an engineering organization suffers from: - A toxic meeting culture and chronic context-switching - Persistent unplanned work, on-call interruptions, and production outages - Lengthy build and test wait times, local dev environment toil, and slow deployment pipelines

Winning organizations avoid viewing AI merely as an autocomplete helper for individual developers sitting at their desks. Instead, they position DevX at the center of their strategy and turn AI inward onto system bottlenecks—using it to shorten CI wait times, reduce environment setup toil, and eliminate meeting overhead across end-to-end value streams.

[26:39](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1599s) Overcoming Organizational Barriers and AI Readiness Models

Referencing research from MIT on barriers to enterprise AI adoption, Laura Tacho pointed out that the obstacles stalling companies are rarely technical. They do not stem from the underlying LLM models or the client software wrapping them, but from human and operational failures: - Weak change management and ambiguous user expectations - Poor developer UX across internal workflows - A lack of genuine executive sponsorship, typified by leadership teams mandating AI adoption while never having opened a laptop to run Windsurf, Claude Code, or GitHub Copilot themselves

To diagnose organizational maturity and build a data-backed case for leadership, she directed teams to two established AI readiness frameworks:

- **DORA AI Capabilities Model**: Maintained by the DORA team at Google Cloud (led by Nathen Harvey), this framework goes beyond the traditional four DORA metrics. It leverages empirical data across thousands of engineering teams to correlate specific organizational practices—such as maintaining a clear, well-communicated AI stance—with measurably superior business outcomes (available at `dora.dev`). - **Thoughtworks Forest Framework**: Detailed in white papers at `thoughtworks.com`, this research-backed maturity model assists companies in conducting rigorous internal audits to verify whether their architectural and operational baselines are prepared to capture actual value from AI experimentation.

She concluded by noting that top-performing organizations anchor their AI pilots not in technology tourism, but in solving real, tangible customer problems.

[28:39](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1719s) Grounding AI Exploration in Real Customer Problems

Laura Tacho brings the conversation back to the reality of business constraints by revisiting the space race metaphor. While shooting for the stars sounds exciting on paper, having an entire engineering organization chasing far-flung, speculative goals is financially and operationally reckless. As Laura Tacho warns: “Exploration and going to Mars is great, but that is not sustainable for your whole entire organization to be experimenting with going to Mars.” Chasing moonshots across the board drains financial resources, distracts engineers from core business objectives, and fails to deliver actual value to customers.

Instead, companies must deliberately split their efforts. Experimentation should absolutely continue, but it needs to be sharply directed toward resolving pressing, real-world customer pain points. By channeling technical exploration into solving immediate organizational and user challenges rather than indulging in unguided science experiments, companies can convert the promise of AI into measurable business performance.

- Broad, unguided exploration across an entire company creates massive expenses without customer payoff. - High-impact AI experimentation must remain laser-focused on existing customer problems to drive tangible organizational outcomes.

---

[29:03](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1743s) Closing: Balancing Moonshots with Pragmatism

Looking toward the horizon, Laura Tacho invokes Carl Sagan's classic ethos to capture the current technological inflection point: “Somewhere something incredible is waiting to be known.” The rapid emergence of generative models and autonomous agents is dramatically expanding the boundaries of software engineering, reshaping who software can be built for, what can be created, and the velocity at which teams can deliver. The industry has clearly entered a new age of technological exploration, filled with immense potential.

Yet with all of that awe comes a vital responsibility to remain anchored in practical execution. Laura Tacho closes her keynote by urging the audience to balance ambitious vision with operational discipline as they attend the rest of the summit. As Laura Tacho emphasizes: “The thing I want to urge all of you to take with you into the rest of the sessions today is to find that balance between a sense of wonder and a sense of awe and aiming for Mars and aiming for your moon colony but also understanding that we need to solve the problems here on Earth and we have to live in this reality.” She leaves engineering leaders with a final piece of guidance for navigating the AI era: stay grounded, maintain a healthy skepticism, preserve human judgment, and above all, remain pragmatic.

```mermaid flowchart LR accTitle: Balancing AI Vision with Pragmatic Execution accDescr: A diagram contrasting ambitious long-term AI exploration with grounded, customer-focused daily execution. subgraph Wonder ["The Moonshot Vision"] A["Sense of Awe & Exploration"] --> B["Aiming for Mars & Moon Colonies"] end subgraph Grounded ["Everyday Execution"] C["Solving Problems on Earth"] --> D["Laser Focus on Customer Needs"] end Wonder <-->|"Maintain Healthy Balance"| Grounded ```

- Generative models and autonomous agents are accelerating technical possibilities and broadening the scope of software delivery. - Engineering organizations must strike a conscious balance between awe-inspiring exploration and solving day-to-day problems in current reality. - The closing mandate for leaders navigating AI hype: stay grounded, stay skeptical, stay human, and stay pragmatic.

## Notable quotes > [20:14](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1214s) Organizations are constrained by human and systems level problems. We remain skeptical of the promise of any technology to improve organizational performance without first addressing those human and systems level constraints. — Laura Tacho

> [09:54](https://www.youtube.com/watch?v=LOHgRw43fFk&t=594s) If we want organizational results we have to think about it on an organizational level not on a coding task level. — Laura Tacho

> [19:25](https://www.youtube.com/watch?v=LOHgRw43fFk&t=1165s) I believe consensus among agents is going to be a huge problem to solve in 2026. — Laura Tacho

> [01:14](https://www.youtube.com/watch?v=LOHgRw43fFk&t=74s) We need to keep the sense of wonder while also balancing it with pragmatism and beating the hype by looking at data. — Laura Tacho

> [07:45](https://www.youtube.com/watch?v=LOHgRw43fFk&t=465s) One thing that is absolutely true, one thing that is common is that there is no typical experience with AI. — Laura Tacho

## Mentioned Laura Tacho, Carl Sagan, Cursor, GitHub Copilot, Claude, Claude Code, OpenAI Codex, ChatGPT, Google, Brian Houck, Microsoft, SPACE framework, MIT, The GenAI Divide, DORA, Martin Fowler, Kent Beck, Gas Town, Moltbot, Clawbot, OpenClaw, GPT-5.3 Codex, OpenAI, Haven Headache and Migraine Center, Linear, Figma, Cisco, JPMorgan Chase, Multi-Agent Framework for Annotation (MAFA), Steve Yegge, Gergely Orosz, Thoughtworks, Abi Noda, DX, Core 4 framework, Windsurf, Nathen Harvey, Thoughtworks Forest Framework

## Video description With Laura Tacho, CTO. Keynote at The Pragmatic Summit

More details in this accompanying article: https://newsletter.pragmaticengineer.com/p/the-future-of-software-engineering-with-ai

The Pragmatic Summit: https://www.pragmaticsummit.com

Timestamps:

0:00 Intro 1:14 The space race analogy for the AI era 3:24 New industry benchmarks in 2026 6:06 Onboarding speed 7:45 Why averages mislead 10:07 High adoption, low transformation 15:22 Codex signals 16:55 Case studies: Haven, Cisco, JPMorgan 20:14 AI won’t fix itself 22:12 What winning orgs do: 3 things 27:19 AI readiness models 29:03 Closing

[Source](https://www.youtube.com/watch?v=LOHgRw43fFk)

Reading summary

Laura Tacho argues that while AI coding assistants have reached near-ubiquitous adoption, they function as organizational amplifiers rather than equalizers, widening the performance gap between mature engineering cultures and dysfunctional ones. Adoption Benchmarks and Developer Velocity Gains (3:24) Benchmarks collected between November 2025 and February 2026 across 121,000 developers at more than 450 companies indicate that AI adoption has largely plateaued: 92.6% use coding assistants monthly and approximately 75% use them weekly. Developer time savings hold steady at an average of 4.08 hours per week (around a 10% gain), closely aligning with internal productivity lifts reported by Google. However, direct code authorship is accelerating sharply. Across 42,600 tracked developers, 26.9% of upstream merged code was authored directly by AI without significant human intervention, rising to over 30% among daily active users. Operational acceleration is most pronounced in developer onboarding. Cross-industry data reveals that the time required for a developer to ship their tenth pull request dropped by 50% between Q1 2024 and Q4 2025. Citing Microsoft research by Brian Houck, Tacho notes that the velocity achieved while ramping up to that tenth pull request establishes a persistent performance trajectory that endures for an engineer's first two years of employment. The Amplifier Effect and Quality Divergence (7:45) Tacho cautions that aggregate industry averages obscure a stark operational divergence across organizations. Invoking the Big Bang analogy, she emphasizes that AI does not level organizational disparities; it acts as an amplifier that magnifies existing baseline hygiene. Incident telemetry from over 67,000 developers gathered between late 2025 and early 2026 shows that organizations with disciplined testing frameworks and mature continuous integration halved their customer-facing outages. Conversely, organizations with fragile architectures doubled their customer incident rate, using AI only to deploy unvetted code and compound technical debt faster. This polarization explains why widespread tool adoption rarely correlates with commercial success. Citing an MIT study of 152 organizations titled The GenAI Divide, Tacho points out that companies consistently run into a productivity ceiling when they treat AI solely as an individual, desk-level coding assistant rather than an organizational change management challenge. Emerging Agentic Workflows and Coordination Hurdles (15:22) Telemetry from pioneering enterprises indicates that software delivery is rapidly shifting from passive autocompletion to autonomous execution. In a cohort of 3,000 developers across six early-adopting firms, 80% interact with agentic workflows weekly and over 50% rely on them daily. Following the February 2026 release of the Codex desktop app, OpenAI's internal developers using the platform shipped 60% more pull requests per week than peers relying on older tools. Enterprise case studies reflect this shift: Haven Healthcare uses autonomous Ralph loops to convert product requirements into tested prototypes, while Cisco cut code review cycles by 50% across 18,000 engineers using Codex. As organizations build specialized, multi-agent pipelines—where separate tiers of agents ingest, annotate, calibrate, and validate work—systemic orchestration becomes essential. Tacho forecasts that establishing formal consensus algorithms among conflicting, independent agents will emerge as a defining technical challenge for software engineering teams in 2026. Technical Hygiene as the Prerequisite for ROI (24:04) Reflecting on discussions at the Agile 25th anniversary retreat with Martin Fowler, Kent Beck, and Steve Yegge, Tacho warns that artificial intelligence cannot repair systemic dysfunctions; unaddressed organizational flaws will simply be carried into automated workflows. The modest four hours saved weekly by faster code entry is quickly negated if an organization suffers from chronic context-switching, excessive meetings, and slow build pipelines. Foundational developer experience practices—comprehensive documentation, clean service boundaries, and fast feedback loops—are the exact operational prerequisites required for autonomous agents to function effectively. To capture actual value, leaders must abandon indiscriminate "spray and pray" seat licensing in favor of rigorous measurement. Tacho points to frameworks such as the AI Measurement Framework (co-authored with Abi Noda), the DORA AI Capabilities Model, and Thoughtworks' Forest Framework to evaluate prompt usage, delivery speed, code quality, and compute costs against actual customer outcomes. Tacho concludes that organizations must counter industry hype by balancing exploration with operational pragmatism. Long-term commercial success requires engineering leaders to direct agentic capabilities toward resolving immediate customer problems and internal delivery bottlenecks rather than pursuing broad, unguided technical moonshots.

Source quotes

[20:14] Laura Tacho: Organizations are constrained by human and systems level problems. We remain skeptical of the promise of any technology to improve organizational performance without first addressing those human and systems level constraints. [9:54] Laura Tacho: If we want organizational results we have to think about it on an organizational level not on a coding task level. [19:25] Laura Tacho: I believe consensus among agents is going to be a huge problem to solve in 2026. [1:14] Laura Tacho: We need to keep the sense of wonder while also balancing it with pragmatism and beating the hype by looking at data. [7:45] Laura Tacho: One thing that is absolutely true, one thing that is common is that there is no typical experience with AI.

Original transcript

[0:04] Today I wanted to have a really [music] [0:06] pragmatic and downto-earth conversation [0:09] about AI, what is actually happening in [0:12] our organizations, what you can expect [0:14] to happen um and how agents are changing [0:16] the game. And I thought in order to have [0:18] this really pragmatic down-to-earth [0:20] conversation, I wanted to take us to [0:23] space. [0:25] I do see a lot of parallels between the [0:27] age of exploration and the space race [0:29] and the age of AI. So last week I was [0:32] talking with a CTO co-founder of a small [0:34] startup. I was also talking with a [0:36] principal engineering lead at a very big [0:38] bank, highly regulated and we sat for a [0:41] solid 15 minutes talking about all the [0:43] cool stuff that we were building, how [0:45] it's brought back joy, the joy of [0:46] coding, and we just had all of these [0:49] ideas and it seems like we couldn't [0:50] build fast enough. There is so much to [0:54] learn and so much to build. And it [0:56] reminds me of this really lovely quote [0:58] from Carl Sean that I love that [1:00] somewhere something incredible is [1:02] waiting to be known. And I think that [1:04] this quote really captures what a lot of [1:06] us are feeling about AI and about the [1:09] experimentation and just the possibility [1:10] that is out there. [1:14] This is the same feeling that we had [1:15] with the age of space exploration, going [1:17] to the moon, going to Mars. But it [1:20] didn't come without skepticism. [1:22] Why spend all of this money [1:24] experimenting and going to the moon when [1:26] we had lots of problems to solve here on [1:28] Earth? [1:30] We had a lot of wonder, but we also had [1:32] a lot of skepticism because space [1:34] exploration wasn't just about science. [1:36] It was also global. It was economic. It [1:38] was political. [1:40] Space wasn't a silver bullet to solve [1:42] all of the problems that we had with [1:43] humanity. But we also can't deny that [1:46] when a man landed on the moon that it [1:48] was a very pivotal defining moment for [1:51] all of humankind and had a sense of [1:53] wonder and had the world in awe. It was [1:55] about redefining what was possible. [1:59] And similarly, we have a lot of wonder [2:01] and a lot of optimism and a lot of [2:03] promise about AI. We can talk about [2:05] productivity boosts and all of the hype [2:07] around, you know, 100% productivity [2:09] boost and all of our code being written [2:11] by AI. We have the promise of perhaps [2:13] the first singleperson billion-dollar [2:15] startup with a person with an idea and [2:18] an army of agents. There's a lot of [2:20] optimism out there. Similarly, there's a [2:23] lot of skepticism. There's a lot of [2:25] skepticism in the corporate world about [2:27] the real economic impact of AI given how [2:30] expensive it is, given the environmental [2:31] impact. There's also a lot of skepticism [2:34] in a lot of different studies about the [2:35] real productivity impact. In certain [2:38] circumstances, it can be really um it [2:40] can really accelerate and in other [2:42] circumstances it can actually slow us [2:44] down and get in our way. It's hard to [2:45] know what's real. [2:48] But as technology changes, it is really [2:50] good and fine to have that sense of [2:52] wonder of exploring the universe while [2:55] also realizing that we have problems [2:56] here on Earth to solve. We have to learn [2:58] how to balance that sense of wonder and [3:00] curiosity with the acknowledgment that [3:04] we are living in reality and we need to [3:06] keep our our feet firmly planted on this [3:08] earth. We need to understand how these [3:10] experiments are actually going to apply [3:11] to everyday companies. How are we [3:13] actually going to improve the world [3:14] around us? We need to keep the sense of [3:17] wonder while also balancing it with [3:19] pragmatism and beating the hype by [3:22] looking at data. And so that's what I [3:24] want to do right now. I'm going to share [3:26] some brand new AI industry benchmarks [3:28] with you. This is new data that no one [3:30] has ever seen ever before in the world. [3:32] Okay, it's coming right now to you. Um, [3:34] I just pulled these down. I guess the [3:36] the static team has seen them because [3:37] they they saw the preview of my slides, [3:39] but aside from them, no one has ever [3:40] seen it. Um, this is though not really [3:43] surprising because a lot of these [3:44] numbers have not changed very much from [3:47] the last quarter. So, what we're looking [3:48] at here is a sample of 121,000 [3:50] developers at over 450 companies. This [3:53] data was pulled from November through [3:55] February 1st, 2026. I I really just did [3:58] this. Um, we're sitting around 92.6 of [4:01] developers are using an AI coding [4:03] assistant at least once a month to get [4:04] their work done. And about 75% of [4:06] developers are using an AI coding [4:08] assistant at least once a week. When I [4:11] say a AI coding assistant, most [4:13] developers define that as cursor, Codex, [4:16] Copi, you know, Claude, not ChatgPT [4:18] necessarily. Um, but it is a bit [4:20] open-ended, so keep that in mind. [4:24] When it comes to time savings, time [4:26] savings is not the only measure of [4:29] productivity impact, but it is an [4:31] important signal. It's a good leading [4:32] indicator. We're sitting around 4.08 [4:36] uh self-reported hours saved due to AI [4:38] tool usage per week per developer. This [4:41] is not all too different from the number [4:43] that came in Q2 of 2025. And then the [4:47] number for Q4 of 2025 was about 3 uh six [4:51] or seven. So this is kind of hovering [4:54] around the 4 hour mark. And there's been [4:56] a few articles, for example, from Google [4:57] in the last year citing about a 10% [5:00] productivity increase. And if we look at [5:01] it in terms of time savings, we're kind [5:03] of hovering around that 10% mark. It [5:06] hasn't changed dramatically um over the [5:09] last few quarters. [5:11] What is changing and what is moving up [5:14] very quickly is the amount of code [5:17] getting merged upstream or in a [5:19] customer-f facing environment that was [5:21] written by AI that was merged without [5:23] significant human intervention. We call [5:26] that AI authored code. And in a sample [5:28] of around 42,600 [5:30] developers from that same time frame, [5:33] uh, November 1st to February 1st, 2026, [5:36] we're at about 26.9% industrywide for [5:40] all of these developers. That's how much [5:41] code is hitting production that was AI [5:43] authored. [5:45] This is moving up from 22% in the last [5:47] quarter, which is actually a pretty [5:49] significant change quarter over quarter. [5:51] And we can see that daily users of AI [5:53] have crested over that 30% mark. So [5:55] almost a third of their code is being [5:57] written by AI that is actually being [5:59] merged, passing through code review and [6:01] getting into a customer-f facing [6:03] environment. [6:06] One of my favorite use cases for [6:08] applying AI is to onboarding. And I had [6:10] a bit of a hunch that AI was going to be [6:12] a great tool for onboarding, helping [6:14] connect people with information earlier [6:17] and sooner. And I have all of this data [6:20] and I thought, let me look at this [6:21] quarter over quarter. And in fact, if we [6:24] look at Q1 of 2024 all the way over here [6:26] on the left side and fast forward to Q4 [6:28] of 2025, we have about a half uh we have [6:31] a half reduction in onboarding time. [6:34] This is looking at the time to 10th PR. [6:36] So by the time a developer hits their [6:38] 10th PR, that's a pretty important [6:39] onboarding milestone that the industry [6:41] has mostly aligned on in terms of [6:43] onboarding and that has been cut in half [6:47] now. And when we correlate that with the [6:49] uptick of AI usage, it makes a really [6:51] pretty graph. Uh AI is fantastic for [6:55] onboarding. And this is not just brand [6:57] new hires to your company. We've also [6:59] seen plenty of evidence that this is for [7:02] engineers who are moving projects or [7:03] even non-engineers coming onboarding [7:06] into projects. What's really important [7:08] about this number is that there was a [7:10] separate study done um by Brian Hulcat [7:12] at Microsoft. He's the co-author of the [7:14] space framework of developer [7:15] productivity and they found in [7:17] Microsoft's context that the time to [7:19] 10th PR actually that performance sticks [7:21] with an engineer for their first two [7:23] years of tenure. So if you onboard [7:25] faster that productivity gain isn't just [7:27] onboarding it actually sticks with them [7:29] for at least two years after they have [7:31] started at the company. So this is a [7:33] very important and significant uh trend [7:36] that we're seeing here with using AI to [7:38] connect developers, reduce cognitive [7:40] load, and get them onboarded more [7:42] quickly into their code bases. [7:45] One thing that's really important for me [7:47] to call out, although I have just shared [7:48] with you industry benchmarks, averages [7:51] are just math. And as the polls move [7:54] further away from each other, the [7:55] average stays the same. Average does not [7:57] mean typical. It does not mean what is [7:59] going to happen to you. And it doesn't [8:01] mean what a common experience is. One [8:04] thing that is absolutely true, one thing [8:06] that is common is that there is no [8:07] typical experience with AI. [8:10] There is no typical experience with AI. [8:12] It is it is extremely different in every [8:15] single company because every company has [8:17] their own problems and their own [8:18] culture. [8:20] This uneven impact can take us back to [8:23] space for just a minute. [8:26] So we can go back to the origins of the [8:28] universe. We had the big bang and there [8:30] was this massive release of energy and [8:32] as this energy released the time the the [8:34] space and time in between objects grows [8:36] bigger right things are moving apart and [8:39] for a lot of us the emergence of AI and [8:42] AI coming into our organizations and in [8:44] the industry has felt a lot like this [8:46] big bang we've had this explosive [8:48] release of energy in the center of our [8:49] world and things keep moving apart [8:52] organizational performance is [8:54] multi-dimensional and these [8:55] organizations are just going off into [8:57] different extremes teams based on what [8:59] they were doing before. AI is an [9:01] accelerator. It's a multiplier and it is [9:03] moving organizations off in different [9:05] directions. The best example I can share [9:08] with you of this is quality. Okay, so in [9:13] this case, this is not every [9:15] organization, but some organizations are [9:18] facing twice as many customer-f facing [9:20] incidents and this is from a sample of [9:22] over 67,000 developers from Q1. So that [9:25] same time frame of uh November to [9:27] February. So just looking in that time [9:28] frame organizations are experiencing [9:30] twice as many customerf facing incidents [9:33] at the same time at the same time [9:37] companies are also experiencing 50% [9:39] fewer incidents. So some companies have [9:42] used AI they have a really healthy [9:44] system it has amplified that system they [9:46] are seeing fewer incidents they're [9:48] moving faster they are accelerating with [9:49] higher quality higher code [9:51] maintainability higher change [9:52] confidence. On the other side though, [9:54] blasting off into the other part of the [9:56] universe, we have organizations who were [9:58] dysfunctional already. No, they're more [10:00] dysfunctional. They're dysfunctional and [10:02] dysfunctional faster. [10:05] Okay. [10:07] Similarly to this uneven impact, [10:10] organizations are seeing really uneven [10:12] results like economically from using AI. [10:16] there are a lot of steep drop off drop [10:18] offs when it comes to using AI in a [10:19] pilot context to production and then [10:21] actually trying to tie it to profit. [10:24] This is from uh an MIT study that was [10:26] published in July of 2025 called the Gen [10:29] AI divide. And what the study concluded [10:32] they did a survey of 152 organizations [10:34] was that right now where we are in the [10:36] industry is that we have really high [10:37] adoption, right? That 92.6 number. Um [10:40] DORA also does its own research. We're [10:42] hovering around that 90% adoption [10:43] number. high adoption but actually low [10:46] transformation because as it turns out [10:49] transformation is really uncomfortable [10:51] and organizations that were ready to [10:53] give up on the cloud transformation on [10:56] the agile transformation are also giving [10:58] up on their AI transformations. It is [11:01] really really difficult to look at your [11:04] whole organization and look at the [11:05] problems and think we got to change [11:08] something about this and that is what [11:10] organizations need to do in order to [11:11] actually see change to their bottom [11:14] line. [11:16] All of this to say back to my previous [11:18] point we have 92.6 adop uh percent [11:20] adoption among developers in our [11:22] industry but adoption doesn't mean [11:24] impact. Using the tool doesn't mean that [11:26] it's going to actually advance your [11:28] organization or do anything. It is an [11:31] organizational problem that needs [11:32] organizational change management. But [11:35] that's not really what we were promised [11:36] with all of the hype was like, hey, [11:38] experiment with AI and then something [11:39] happens and then we profit. [11:42] What happens though is that these tools [11:44] were primarily deployed into individual [11:47] coding tasks. And what this MIT study [11:50] found in this high adoption low [11:51] transformation is that when we apply it [11:53] only to the surface area of a developer [11:55] sitting at their desk there is a very [11:58] very low ceiling of productivity gain. [12:01] This is an organizational problem. If we [12:03] want organizational results we have to [12:04] think about it on an organizational [12:06] level not on a coding task level. [12:10] Fortunately, our universe is expanding [12:11] right now [12:13] and that expanding is coming through the [12:16] use of agents in agentic workflows. [12:19] Our universe is getting bigger and so [12:21] are all of the promises and all of the [12:24] hype, but so is the possibility. [12:28] So, let's go back to the moon landing, [12:30] right? Like the ultimate hype was that [12:32] we're all going to be living on the moon [12:33] by now in flying cars like jetson style. [12:37] Um, similarly here we have a little bit [12:40] of like crazy ideas. Um, Gas Town, if [12:42] any of you have used it, um, there [12:44] there's just like there's so much crazy [12:46] stuff to do right now. Um, Gas Town is [12:48] infinitely interesting to me. There are [12:51] so many interesting things. Um, [12:52] disclaimer, don't use Gas Town. It is [12:54] unhinged. Um, we've got OpenClaw, [12:58] Maltbot, Clawbot, whatever it's called. [12:59] We've got Ralph loops. We've got all the [13:01] stuff, right? There is so much [13:02] experimentation and so much fun. It's [13:05] just really fun to build. Um, but me [13:08] building my nail polish matching like [13:11] color scheme app while I'm sitting at [13:12] the nail salon is not the same as a [13:15] multinational bank being able to change [13:17] their revenue because of AI. Those are [13:19] really different things. Um, and I was [13:22] at this retreat with Martin Fowler and [13:25] Kent who are I think back there. Hello. [13:26] We'll talk about that a bit more later. [13:28] We spent a lot of time trying to connect [13:30] AI and the use of AI to bottom line, to [13:33] profit, to P&L. And interestingly, kind [13:35] of where we landed at the end was this [13:37] question of like what is the value of [13:39] innovation? Was it still valuable to go [13:41] to the moon even though I'm not really [13:43] located on the moon right now? And I [13:45] would argue that yes, it is valuable to [13:48] innovate. And that can get into some [13:50] murky area because this is a business, [13:52] right? This isn't just society and and [13:55] doing things for the good of humankind. [13:56] we have to do them in an economic [13:58] context and that can get a little bit [13:59] tricky. [14:01] So when we think about this quote [14:03] something uh somewhere something [14:04] incredible is waiting to be known. There [14:07] is a sense of wonder and AI and space [14:09] are both the age of exploration and it [14:12] is so exciting. [14:15] But the point of going to the moon [14:17] wasn't that we all need to live on the [14:19] moon. In fact, the point of going to the [14:21] moon and the point of exploring and [14:22] doing all this crazy stuff was to [14:24] improve life on Earth. It was to use the [14:27] space exploration and all of this wonder [14:29] to apply it to the systems level [14:30] problems that we had back on Earth. Not [14:33] everyone wants to live on the moon. Um, [14:35] but we have sunglasses, we have space [14:37] blankets, we have barcodes, we have [14:39] quartz watches. We have so much [14:41] technology and so many improvements back [14:44] on Earth because of this crazy age of [14:46] exploration where we all went to space. [14:49] Even though we're not living on the [14:50] moon, we've still used the lessons and [14:52] applied it to our systems back here on [14:54] Earth. [14:55] And so thinking about agentic workflows, [14:57] agents expand the possibilities of what [15:00] we can build, how we can build it, and [15:02] who we can build it for. Not everyone [15:04] goes to the moon, and it's okay not to [15:06] go to the moon. Not everyone is going to [15:08] be building crazy stuff with Gas Town [15:11] every day in in your enterprise context. [15:14] And that's also okay because the [15:15] experimentation helps push the boundary [15:18] of what's possible and helps us think [15:19] about solving problems in new ways. [15:22] So, let's talk a little bit about how [15:24] agents are being used in the industry [15:25] right now. Again, this is new data that [15:27] I'm sharing for the first time here. Um, [15:29] agentic use is on the rise. There's not [15:31] a lot of companies, honestly, that are [15:33] so far ahead of the curve that they're [15:34] already uh instrumenting their agentic [15:37] use cases with really good telemetry. [15:38] This sample is a little bit smaller. [15:40] It's around 3,000 developers at six [15:42] companies. Keep in mind, these companies [15:44] are ahead of the curve. They're already [15:45] instrumenting their agentic workflows [15:47] with telemetry. Um, we have about 80% of [15:51] developers using these agentic workflows [15:53] at least once a week with over 50% using [15:56] agentic workflows every single day to [15:58] get their work done. [16:00] We talked about codecs I think in the [16:02] previous panel. So on February 2nd, the [16:05] Codeex desktop app was released and [16:07] since then there's been a million over a [16:09] million downloads by now. I got this [16:10] data yesterday. I'm sure it's quite [16:12] different by now. There's been a 60% [16:14] growth in users just in the last week. [16:17] Um they also launched uh GPT 5.3 codecs [16:21] uh last Thursday. They're processing [16:23] trillions of tokens per week. Internally [16:27] at OpenAI, 95% of developers are using [16:29] codecs to ship stuff. And of the [16:33] developers who are using codecs versus [16:35] other AI tools, the developers who use [16:37] codecs are shipping about 60% more PRs [16:40] per week, which is very interesting. a [16:42] data point, not the only data point, but [16:44] it just speaks to the very high ceiling, [16:47] the high possibility, the sense of [16:48] wonder that we have with building all of [16:50] the stuff with cool new tools like [16:52] Agentic Workflows. [16:55] I want to bring it back to a nonAI [16:58] startup though. Um, I want to highlight [16:59] Haven Headache and Migraine Center. So, [17:01] this is a company that's based here in [17:03] San Francisco, actually just a few [17:04] blocks away. Haven set out to answer the [17:07] question, can we solve headaches with [17:09] Zoom? And it turns out you can. Um, so [17:12] if you're a headache sufferer, this [17:13] might be useful for you to to learn [17:14] about. In healthcare, it's really really [17:17] uh crucial for Haven and their [17:19] development team to distinguish between [17:21] using agents for durable code or [17:23] disposable code. One of the things that [17:25] they're doing that's very cool since [17:26] they are a disruptor, they are a small [17:28] startup is using um using agentic [17:30] workflows to rapidly prototype new [17:32] custom uh like new patient workflows. So [17:34] they're working on a patient portal [17:36] building with Ralph loops taking uh [17:38] linear and Figma artifacts changing it [17:40] into a PRD uh you know spitting that out [17:43] in JSON and then just having Ralph loops [17:45] run. What they're getting though isn't [17:47] garbage disposable AI slop. What they're [17:49] getting is really high quality [17:51] prototypes with really excellent [17:53] documentation, excellent tests, much [17:55] higher quality at a way faster rate than [17:58] they would have um if they would have [17:59] built it by hand the oldfashioned way. [18:03] The other thing that they're doing that [18:04] I really admire is improving the [18:06] standard of care for their patients by [18:08] training a HIPACO compliant model on [18:09] hundreds of thousands of symptom logs. [18:12] So Haven meets you where you're at. You [18:13] get a text message, you can log your [18:15] symptoms and then they can um instrument [18:17] your care, figure out what needs to [18:18] happen from there. So they're training a [18:20] hypocmplant model on hundreds of [18:21] thousands of these messages so that [18:23] those messages can be routed to, you [18:25] know, medication refill or schedule [18:27] follow-up appointment just meets you [18:28] where you are. And the result of this is [18:30] that they have 3x the industry average [18:33] in customer satisfaction for a [18:34] healthcare tool like this, but also real [18:37] real meaningful clinical outcomes. So [18:39] their patients have fewer headache days [18:41] per month and also the severity of their [18:43] headaches is much uh much less severe. [18:46] So good job Haven. In the enterprise, [18:49] there are lots of examples of big [18:51] enterprise companies experimenting with [18:52] agent workflows. So there's an [18:54] enterprise manufacturing company that's [18:56] using it for solely internal developer [18:58] purposes. They used C-pilot and Claude [19:01] to build out a dev portal to accelerate [19:03] uh developer onboarding. At Cisco, [19:06] there's 18,000 engineers using codecs [19:09] daily. They're using the uh codecs for [19:11] complex migrations and also code review [19:14] leading to a 50% reduction in the amount [19:15] of time it takes to do code review. [19:18] There's a really cool paper as well by [19:19] JP Morgan Chase's multi-agent framework [19:22] for annotation, MAFA. If you Google [19:24] that, you can find the source paper. [19:25] It's really fascinating. Um, what [19:27] they're doing is building out like a [19:29] whole business of agents. So like a true [19:31] multi- aent workflow similar to Gastown [19:33] where each agent has a special job to [19:35] do. What they're also doing in this [19:37] model is introducing consensus among the [19:39] agents. So they're taking all of these [19:42] interactions and then they're annotating [19:43] them. this was you know the intent what [19:45] was it an FAQ what what were all these [19:48] interactions the agents are annotating [19:50] them and then there's another set of [19:52] agents who are responsible for reranking [19:54] and calibrating and validating the [19:56] output and then of course we have to [19:58] introduce consensus algorithms to the [20:00] party because now we have multiple [20:02] agents with maybe multiple different [20:04] opinions about things um this is really [20:06] fascinating and I believe consensus [20:08] among agents is going to be a huge [20:09] problem to solve in 2026 [20:14] I spoke about this retreat. I was lucky [20:16] enough to be invited by Martin Fowler [20:18] and thoughtworks to the future of [20:19] software development retreat celebrating [20:21] the 25th anniversary of the agile [20:23] manifesto of Gerge joined me. A few [20:26] other folks who are here also joined me. [20:28] We spent a day and a half up in the [20:30] mountains talking about agents. That's [20:33] really all we talked about about using [20:34] agents responsibly, ethically, [20:36] sustainably, how we can use them for [20:38] organizations. And our conclusion even [20:40] though there was so much interesting [20:42] stuff, Steve Yaggi was there, we were [20:44] working on Gas Town things, like there [20:46] was a lot of experimentation happening, [20:48] but the conclusion that we came to was [20:50] that AI does not solve organizational [20:52] systems problems. It only can do that [20:54] when you apply AI to the system problem, [20:56] which means you need to acknowledge that [20:58] the system problem exists in the first [21:00] place. AI is not a magic silver bullet. [21:02] Even though things like Gastown exist, [21:04] even though there is so much sense of [21:06] curiosity and wonder in the universe, [21:09] um we kind of had a sort of off-the cuff [21:12] conversation. Uh Kent Beck, Steve and I [21:14] were just catching up um outside uh of [21:17] one of the sessions in between [21:18] conversations. And here's sort of where [21:20] we summarized our thoughts. [21:21] Organizations are constrained by human [21:23] and systems level problems. We remain [21:26] skeptical of the promise of any [21:28] technology to improve organizational [21:30] performance without first addressing [21:32] those human and systems level [21:33] constraints. We remain skeptical and we [21:36] also remain human because the risk is if [21:39] we don't address the systems level [21:41] problems, we will just take them to [21:43] space with us. We will just take them to [21:46] space with us. We're not actually going [21:47] to solve the human factors that are the [21:50] driving force behind all of the [21:52] constraints that organizations have [21:53] right now. We can apply AI to those [21:55] problems, but we still need to solve [21:57] them. We can't just go to the moon and [21:59] expect that pollution and garbage and [22:01] traffic aren't going to be a problem [22:02] anymore. [22:05] And so the question is not how to [22:06] colonize Mars, but the question is how [22:08] to get real organizational impact with [22:10] agents and AI. At this retreat, we also [22:13] talked a lot about common factors that [22:15] we see. What do we see organizations [22:17] doing? What are the common patterns that [22:19] is kind of like the secret to to [22:21] winning? What do they have in common? [22:23] The first one is that organizations who [22:25] win with AI and are winning with AI have [22:28] goals and they measure their progress [22:30] against those goals. Spray and prey does [22:33] not work. Spray and prey, what I mean by [22:36] that is just giving all of your [22:38] developers licenses and hoping for the [22:40] best. It does not work. I can say that [22:43] very very clearly. I have a lot of [22:44] evidence that does not work. If you can [22:47] point AI innovation and that [22:49] experimentation to a problem, have a [22:51] concrete goal and then measure if you're [22:53] reaching that goal, that is what winning [22:55] organizations are doing right now. [22:57] Because as Spock has told us, [22:59] insufficient facts always invite danger. [23:01] We need to measure things. We need to [23:03] have data. And I know this is something [23:05] that's really difficult for a lot of [23:06] organizations right now because [23:08] developer productivity and engineering [23:10] excellence are also really hard [23:12] problems. And this is happening all at [23:13] the intersection. So I have something [23:15] that can help you if that is in uh a [23:18] problem that you're facing in your [23:19] organization. This is the AI measurement [23:21] framework. This is a framework that I [23:22] co-authored with Abby notto who's the [23:24] CEO of DX. [23:26] This complements our core four framework [23:28] which some of you might have heard [23:29] otherwise it's in the impact column [23:30] here. What we're looking to do is track [23:32] not just usage and adoption and [23:34] utilization of AI but then also [23:36] translate that into real organizational [23:37] impact. Is this changing your speed, [23:41] your developer experience, your quality, [23:43] your innovation ratio? Those are really [23:45] important questions to connect the [23:47] adoption to impact. Finally, we have to [23:50] look at the cost. Are we getting a good [23:52] deal? [23:54] Maybe some of us are for now. Um, and we [23:57] need to understand as the cost of these [23:58] tools keeps going up and up, is the [24:01] investment the right one? [24:04] The second thing that is helping [24:06] organizations win is that developer [24:09] experience matters now more than ever. [24:12] Here is a piece of very unconventional [24:14] advice that I'll give you is just [24:15] anything that you were going to talk [24:16] about with your leadership team about [24:18] developer experience. Just call it agent [24:20] experience and you'll get money for it. [24:24] It's funny but it works. Um, it works [24:27] because developer experience, feedback [24:29] loops, um, you know, clearly defined [24:31] services, great documentation, fast CI, [24:35] these are all things that we have been [24:36] screaming about for decades, literally, [24:39] and we've been begging for pennies from [24:41] our organizations to please let us [24:43] invest, please let us invest in [24:44] developer experience. And we've been [24:46] told no over and over again. Come to [24:48] find out, in fact, these are the things [24:51] that make AI really successful. We need [24:52] to have really solid testing and quality [24:55] practices. We need to have great [24:56] documentation. These are critical for [24:58] agentic workflows. It is disheartening [25:01] that we didn't want to spend the money [25:02] when it came to human engineers, but [25:04] when it comes to robot engineers, we're [25:06] okay with it. But that is the world that [25:08] we live in and let's capitalize on our [25:09] opportunity. So Devex matters more than [25:12] ever. In fact, when we look at the data [25:13] right now, remember we're hovering [25:15] around that 4hour mark for time savings. [25:18] when we look at all of the other factors [25:20] of developer experience like AI time [25:22] savings is not going to make up for bad [25:24] meeting like bad meeting culture and [25:26] lots of interruptions and um you know [25:29] developers who are constantly being [25:31] pulled out of their work unplanned work [25:33] interruptions outages those kinds of [25:34] things AI will not make up for that we [25:36] can use AI to help solve that problem [25:38] but AI in and of itself is not going to [25:40] make up for it then when we look kind of [25:43] in the bottom half build and test wait [25:45] time toil and dev environment we put all [25:47] that together We realize that just the [25:49] time savings from coding task speed up [25:51] isn't going to get us very far. But what [25:53] will get us far is when we can take AI [25:56] and point it at those problems. Can we [25:59] use AI to help reduce meeting frequency? [26:01] Can we use AI to improve CI weight time? [26:04] Can we use AI to reduce dev environment [26:06] toil? That is what winning organizations [26:08] are doing right now. They are putting [26:10] DevX at the center of their universe and [26:12] seeing AI as a tool to fix systems level [26:15] problems. [26:18] They're doing it also on an [26:19] organizational level. If you want [26:21] organizational outcomes like revenue, [26:25] P&L, time to market, you have to think [26:28] about AI on an as an organizational [26:30] problem, not as an individual problem [26:32] that your developer needs to solve at [26:33] their desk. It has to apply to workflows [26:36] that span entire value streams. [26:39] back to that MIT study when we looked at [26:41] the barriers to organizational adoption [26:43] uh or the organizational barriers to AI [26:45] adoption, they weren't technical. This [26:48] wasn't about the models necessarily. It [26:50] wasn't even about the tools that wrap [26:52] the models. It was about things like [26:54] change management or lack of executive [26:56] sponsorship when you have an executive [26:58] team saying go with AI, but they [27:00] themselves have never cracked their [27:01] laptop open and fired up windsurf or [27:04] cloud code or codeex. um poor user [27:07] experience, just very unclear [27:09] expectations about AI. Those are the [27:11] things that get in the way. [27:13] If this sounds familiar to you and [27:15] perhaps your organization could do a [27:17] better job, there's two things that I [27:18] want to point you to. Um the first one [27:20] is the Dora AI capabilities model. These [27:22] are models that kind of communicate and [27:24] help you get ready for AI. So think [27:27] about this as an AI readiness model or [27:28] an AI capabilities model. This has in a [27:31] crazy amount of data from organizations [27:33] that Dora studies. They do a lot more [27:34] than just the four key door metrics. Um, [27:38] finding correlations between practices [27:40] that organizations have and good [27:42] outcomes with AI. So, if you use AI and [27:44] have a good clear and communicated AI [27:46] stance, you are going to do better [27:48] organizationally than a company that [27:50] does not have one. You can find this at [27:51] dora.dev. It's the Dora AI capabilities [27:53] model. There was just a new paper that [27:55] came out last month. Last month, uh, [27:58] Nathan is here who leads Dora over at [28:00] Google Cloud. If you want to talk to him [28:01] about this, he's probably the the guy. [28:03] Um the other one is the thoughtworks [28:05] forest framework. This is similar to the [28:07] AI capabilities model kind of a [28:08] different flavor on it. Um if you go to [28:10] thoughtworks.com and look in their white [28:11] papers you can read through this but [28:13] these are both really solid [28:14] wellressearched industrybacked AI [28:16] readiness models to help convince your [28:19] leadership team if you need that um or [28:21] just help you do an internal audit of [28:22] are we doing the right things to make [28:24] ourselves ready to you know reap the [28:27] benefits of all this experimentation. [28:31] The last thing is that organizations who [28:33] are doing really well with AI right now [28:35] are experimenting by solving real [28:37] customer problems. Again, space [28:39] exploration and going to Mars is great, [28:41] but that is not sustainable for your [28:43] whole entire organization to be [28:44] experimenting with going to Mars. It [28:47] just costs too much money. It distracts [28:48] too much from the core business problem. [28:50] It does not serve your customers. So, [28:52] keep experimentation going. Other [28:55] experimentation can be really laser [28:57] focused on real customer problems that [28:58] you have. And that is how you're going [28:59] to see the organizational results. [29:03] Somewhere something incredible is [29:05] waiting to be known. [29:07] There is so much possibility of how we [29:09] can build what we can build, who we can [29:11] build it for right now with AI and [29:12] agents are just accelerating this. They [29:14] are expanding our universe. [29:17] We are definitely in an age of [29:19] exploration. The urge uh the thing I [29:22] want to urge all of you to take with you [29:23] into the rest of the sessions today is [29:25] to find that balance between a sense of [29:27] wonder and a sense of awe and aiming for [29:30] Mars and aiming for your moon colony but [29:32] also understanding that we need to solve [29:33] the problems here on Earth and we have [29:35] to live in this reality. So please stay [29:37] grounded, stay skeptical, stay human, [29:40] most of all stay pragmat stay pragmatic. [29:43] Thank you all. [music] [29:44] [applause]