AI Startups Overhaul Agents as Inference Costs and UX Gaps Hit Scale

Startups shipping at scale with Google DeepMind — Google Cloud Tech

Google Cloud Tech

Autonomous AI tools are failing to retain users during long task loops and burning unsustainable compute on generative retries, forcing founders to replace hands-off automation with interactive parallel swarms.

AI-assisted summary of Startups shipping at scale with Google DeepMind — original source. Check the original for full context.

Source published:

Key ideas

  • Startups are moving from thin API wrappers toward autonomous multi-agent swarms that run parallel background tasks while engaging users in higher-leverage architecture and design decisions. (4:51)
  • Long-running autonomous agent 'missions' can span days or weeks—such as a 14-day run consuming nearly 150 billion cached tokens to migrate a 50-million-line codebase—requiring interaction models closer to consulting engagements than IDE copilots. (20:22)
  • Autonomous video generation faces steep unit economics, where critique loops and automated retries consume roughly three minutes of vendor compute for every single minute of finished footage, pushing end-user costs to $30–$60 per clip. (30:29)
  • Enterprise AI adoption is now primarily bottlenecked by an 'enablement or mindset overhang' rather than raw model capabilities, as legacy regional enterprises outpace paralyzed digital natives in shipping autonomously generated code. (27:23)

Detailed notes

## Background Published on YouTube by Google Cloud Tech on June 25, 2026, this Google Cloud Next panel is led by Logan Kilpatrick from Google DeepMind alongside startup founders Michele Catasta, Bianca Rangecroft, and Jay Wu. According to the session description, the discussion explores real-world workflows using Google DeepMind's models, technical trade-offs in multimodality, and how fast iteration speed serves as a startup's primary competitive moat.

## TL;DR The panel examines how startups are evolving past thin API wrappers into sophisticated, autonomous agents across enterprise coding, creative media, and consumer styling. Founders discuss tactical UX breakthroughs to manage latency and user churn, such as parallel agent swarms and gamified, educational wait states. The speakers also confront critical production bottlenecks, including the lack of objective outcome verification oracles, punishing unit economics in agentic video generation, and an organizational 'mindset overhang' that slows enterprise adoption. Finally, the group anticipates the rise of native video comprehension and visual abstractions that will transform software development beyond traditional text editors.

## Atmosphere & style The session has a brisk, candid, and builder-focused energy, blending onstage conference banter with genuine post-mortems from the AI frontier. Logan Kilpatrick serves as an enthusiastic, insider host, playfully teasing upcoming DeepMind features while prompting founders to share honest operational mistakes, from failed 200-minute autonomous runs to soaring compute bills. The dialogue feels like a pragmatic backstage debrief among engineers, stripping away generic industry hype to tackle the messy, expensive realities of deploying autonomous agents into production.

## Bulletin - [04:51](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=291s) Startups are moving from thin API wrappers toward autonomous multi-agent swarms that run parallel background tasks while engaging users in higher-leverage architecture and design decisions. - [20:22](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1222s) Long-running autonomous agent 'missions' can span days or weeks—such as a 14-day run consuming nearly 150 billion cached tokens to migrate a 50-million-line codebase—requiring interaction models closer to consulting engagements than IDE copilots. - [30:29](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1829s) Autonomous video generation faces steep unit economics, where critique loops and automated retries consume roughly three minutes of vendor compute for every single minute of finished footage, pushing end-user costs to $30–$60 per clip. - [27:23](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1643s) Enterprise AI adoption is now primarily bottlenecked by an 'enablement or mindset overhang' rather than raw model capabilities, as legacy regional enterprises outpace paralyzed digital natives in shipping autonomously generated code.

## Key takeaways - [02:10](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=130s) Agent roadmaps evolve rapidly and require rewriting software architectures roughly every three months to keep up with frontier model leaps, contrasting with stable, slow-moving infrastructure layers. - [02:10](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=130s) Building agents for non-technical knowledge workers requires abstracting away traditional IDEs in favor of conversational and orchestrating interfaces. - [04:51](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=291s) Media generation products have evolved beyond prompting wrappers to autonomous agents that conduct research, map narrative arcs, and incorporate human-in-the-loop iterative editing. - [10:12](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=612s) Because modern digital work is underpinned by software, model-agnostic coding harnesses with recursive self-improvement loops inherently expand into general-purpose computer-use agents for non-developers. - [12:38](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=758s) Agent harnesses are experiencing an 'MVC moment' standardizing orchestration loops, shifting the primary product bottleneck in creative tools to injecting and evaluating subjective aesthetic 'taste.' - [15:47](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=947s) Uninterrupted autonomous loops can cause user abandonment; parallelizing background swarms while offering collaborative planning checkpoints preserves user retention. - [18:44](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1124s) Consumer applications can mask heavy model processing latency through interactive education, peer social loops, and gamified tasks that simultaneously generate training data. - [20:22](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1222s) Agent autonomy operates on a spectrum ranging from synchronous minutes-long IDE assistance to multi-day, scoping-intensive migration missions. - [22:35](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1355s) Transitioning from per-token billing to outcome-based contracts is stalled by the absence of definitive verification oracles outside of disciplined, mature test suites. - [25:30](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1530s) As foundational models commoditize custom-trained personalizations, consumer AI moats increasingly hinge on brand design craft, rapid go-to-market execution, and visual differentiation. - [27:23](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1643s) Enterprise competitive advantage in software engineering stems from leadership mindset and organizational change management rather than tool selection. - [27:23](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1643s) High inference costs between $0.10 and $0.50 per second for leading video generation models limit autonomous media workflows to well-funded commercial use cases. - [30:29](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1829s) Autonomous self-correction cycles multiply vendor compute expenses, forcing video startups to absorb or pass along high costs for short media outputs. - [31:32](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1892s) Current multimodal agents struggle with video editing because they evaluate video sequentially frame-by-frame rather than comprehending continuous motion across a timeline. - [33:29](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=2009s) Multimodality enables coding agents to programmatically build media components and allows human developers to manage complex architectures through visual abstractions rather than raw code lines.

## Detailed walkthrough ### [00:13](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=13s) Introduction and Framing the Frontier of AI

Logan Kilpatrick opens the panel by welcoming attendees to Google Cloud Next in Las Vegas, joking about whether folks have had time to gamble alongside attending sessions. Introducing his role, Kilpatrick explains that he leads the developer platform team at Google DeepMind, where his group is charged with building APIs and delivering frontier models like the Gemini suite and Veo into developers' hands. Rather than spending time on standard formal speaker introductions, he points out that the panelists' names are already displayed on screen and jumps immediately into the core topic: how startups are actively deploying agents and software coding tools, and how their underlying product architecture and implementation details have radically transformed over the prior six months.

[02:10](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=130s) Replit’s 10-Year Infrastructure and the Fast-Moving Agent Roadmap

Michele Catasta of Replit explains that Replit committed fully to an agent-only architecture roughly eighteen months earlier. When asked why that choice seemed obvious so early, Michele clarifies that it was driven by conviction rather than crystal-clear proof: “I think it was mostly conviction and not clarity back in those days, because agents were not nearly as powerful as they are today.” Replit made this structural choice deliberately because its target audience consists of knowledge workers rather than traditional software engineers. Placing a non-technical worker in front of an IDE is the wrong product surface. While the original philosophy has remained steady, the scale of what their agent builds has grown from simple prototypes into mission-critical internal tools deployed across enterprise organizations to thousands of daily active employees with strict security and data governance.

Addressing whether this shift required rebuilding everything from scratch, Michele describes Replit as a “10 year overnight success.” Replit spent eight years constructing foundational developer infrastructure and sandboxes. The agent became the connective tissue that orchestrates that existing toolchain. However, maintaining that foundation alongside frontier AI creates severe internal friction between development timelines:

```mermaid flowchart LR accTitle: Replit Dual Roadmap Tension accDescr: Diagram illustrating the conflict between Replit's fast-moving agent roadmap and its slow-moving infrastructure layer.

subgraph FastTrack["Agent Roadmap (Hyper-Iterative)"] direction TB A["Frontier Model Leaps<br/>(Every ~3 Months)"] --> B["Frequent Codebase &<br/>Architecture Refactors"] end

subgraph SlowTrack["Infrastructure Layer (Stable)"] direction TB C["Tooling & Sandboxes<br/>(8+ Years Built)"] --> D["Security, Governance &<br/>User Data Integrity"] end

FastTrack -.->|"Tension & Connective Tissue"| SlowTrack ```

Michele highlights that while the underlying infrastructure layer must move slowly and deliberately to safeguard user data, “The agent roadmap moves at a crazy pace, and we throw away part of the software we architected on a regular basis because frontier models make leaps every three months or so.”

[04:51](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=291s) Moving from API Wrappers to Autonomous Multimodal Agents

Jay Wu explains how agentic capabilities have shifted the paradigm for multimodal media creation products. In the prior generation of AI software, companies functioned largely as thin API wrappers: they would ingest a user's intent, package a prompt around it, and query a model to return an isolated image or audio clip. With modern agentic architectures, systems can understand overarching goals, conduct autonomous background research, map out an entire storyline, and generate finished end-to-end video. By integrating human-in-the-loop feedback mechanisms, the product lets users guide iterative edits instead of forcing them to manage prompt strings. This allows media platforms to unlock genuine workflow value rather than simply acting as repackaged model endpoints.

[06:11](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=371s) Whering's Three-Pronged Strategy for Digital Wardrobes and Virtual Try-On

Bianca Rangecroft recounts founding Whering to build an AI personal styling agent inside consumers' closets. When pitching the concept in 2020, venture capital investors repeatedly dismissed the idea as impossible: “Yeah, I mean, honestly, when we first started pitching this in 2020, like an agent that could be a personal stylist in your closet, a lot of investors were like, this is never going to happen.” The company leaned into that original thesis, executing a three-part strategy to construct an enduring competitive moat:

- **Proprietary Wardrobe Dataset:** Creating a product-led growth flywheel via a digital closet consumer app that logs user clothing items, brands, sizing dimensions, and style preferences to train fine-tuned, niche styling models. - **Detailed Virtual Try-On Capabilities:** Gathering contextual daily outfit selfies to map how fabrics fit human bodies in diverse everyday environments, enabling personalized visualization. - **B2B Retail Personalization:** Monetizing the B2C consumer data flywheel to provide recommendation engines and personalized shopping tools for enterprise e-commerce platforms, timed alongside closing Whering's seed funding round.

[08:38](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=518s) Factory's Droids and Enterprise Autonomous Lifecycle

Speaking on the origin of Factory's terminology, the speaker clarifies that they refer to their autonomous systems as "droids" rather than agents. The company launched three years prior, immediately following the debut of GPT-3.5, aiming to automate tasks across the entire software development lifecycle (SDLC) while providing governance and visibility for enterprise clients. The speaker notes that “agent” was largely confined to academic cognitive science literature at the time, prompting them to adopt the name droids.

Building fully autonomous lifecycle droids required navigating an extended period where model reasoning was inadequate: “And it just didn't work for 2 and 1/2 years.” While the high-level seed pitch deck remains identical to what enterprise customers readily buy today, the tactical integration and runtime abstractions underneath were rewritten repeatedly as underlying frontier model reasoning matured. Kilpatrick quips that Disney might eventually take issue with the "droid" branding as the panel turns toward implementation choices.

[10:12](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=612s) Software as the General-Purpose Agent Substrate

Kicking off with some lighthearted banter about naming rights and legal boundaries, the discussion touches on the moniker “Droid.” The speaker jokes that as long as they stay outside of physical merchandising and avoid the lawyers in the room, the name remains fair game for now. Logan Kilpatrick shifts the conversation toward a broader product strategy question: as coding evolves into a general-purpose agent harness where “it’s code all the way down,” he asks whether building software agents is quietly transforming into building universal computer-use agents for non-developers.

The response frames this shift not as an expansion into unrelated business verticals, but as the natural consequence of computer architecture. Because software underpins every single action performed on a digital screen, a local-first, model-agnostic coding agent equipped with self-recursive improvement loops inherently functions as a general-purpose agent for almost any knowledge work. The team observed this internally: their third most active user is an enterprise account executive using Droid to pull data and sell the platform itself. Rather than pivoting into diverse functional domains, their thesis is that the rest of the enterprise is simply being converted into code.

As the speaker summarizes the dynamic: “And I think that means that there's a lot of opportunities, but we don't see it as us moving into other spaces. Instead, we see it as other spaces becoming software.”

- Droid is designed as an extremely competent, local-first, model-agnostic coding harness capable of recursive self-improvement. - Enterprise adoption extends beyond core engineers to product managers, data scientists, product marketers, and sales reps querying datasets directly. - The driving thesis aligns with Marc Andreessen’s classic adage that software is eating the world, meaning mastering code equates to mastering computer operation.

---

[12:38](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=758s) The Harness Evolution and the Challenge of Measuring Taste

Logan turns to Jay Wu to explore the friction between increasingly autonomous, long-running agent loops and the subjective human element required in creative outputs. Jay expresses enthusiasm for modern agent harnesses, drawing a direct historical comparison to web development: before frameworks arrived, engineers had to hand-roll custom ReAct loops and context-engineering routines, just as early web developers built everything from scratch before Ruby on Rails and Django standardized the MVC pattern. While harnesses now solve orchestration reliability, the core product bottleneck has shifted to “taste.”

```mermaid flowchart TD accTitle: Video Generation Agent Taste Layer accDescr: Pipeline showing how raw foundation model generation is refined with taste injection and human rubric evaluation.

subgraph Base ["Foundational Layer"] A["Foundational Multimodal Generators"] B["Homogeneous / Generic Output"] A --> B end

subgraph TasteLayer ["Product Differentiation"] C["Agent Harness (Standardized ReAct Loop)"] D["Taste Injection: Storytelling & Visual Curations"] B --> C C --> D end

subgraph Eval ["Quality Control"] E["Defect Detection (Automated)"] F["Rubric Engineering & Human Eval (Taste)"] G["Social Media Ready Video"] D --> E D --> F E --> G F --> G end ```

In the cutthroat landscape of social media, foundation models left to their own devices produce flat, homogeneous slop that fails to capture viewership. Jay emphasizes that generating compelling videos requires infusing high-level visual judgment, pacing, and storytelling directly into the agent’s skill sets and prompt templates. The central technical hurdle is evaluating that output: while technical glitches and video defects are straightforward to spot programmatically, subjective aesthetic quality cannot be easily automated. Without rigorous rubric engineering paired with human evaluation, shipping updates becomes risky guesswork that can degrade the user experience.

Jay explains the fundamental evaluation dilemma facing creative agents: “And it has always been a challenge in this because it's easier to detect defects in videos, but it's hard to measure taste.”

- Agent harnesses are experiencing an “MVC moment,” standardizing context engineering and ReAct loops similarly to Rails and Django in the 2000s. - Foundation models default to generic aesthetic choices unless proprietary taste layers, curated templates, and storytelling structures are explicitly injected. - Defect detection can be programmatic, but measuring artistic quality requires expensive, continuous rubric engineering and human-in-the-loop validation.

---

[15:47](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=947s) Autonomy Traps and the Parallel Swarm Solution

Logan brings up the tension between speed and accuracy: Google’s classic search thesis proved that sub-100-millisecond response times keep users engaged, whereas waiting tens of seconds leads to severe drop-off. As AI agents tackle deeper tasks that take minutes or hours, builders face an existential UX challenge regarding what the user is supposed to do in the interim. Michele Catasta shares a candid post-mortem from their September 2025 release, where they marketed an agent capable of running autonomously for 200 consecutive minutes. They quickly discovered that uninterrupted autonomy actually backfired because users abandoned the interface entirely.

To fix this disconnect, Michele describes how they overhauled their architecture for their March release. Instead of forcing users to wait on a single monolithic, sequential loop, they introduced an orchestrated swarm of parallel agents managed behind the scenes. While the background swarm executes development workflows, the interface actively engages users in high-leverage planning, UI design, and architectural decisions across the software development lifecycle. By parallelizing the computation and designing collaborative touchpoints, the platform eliminates the dead air that causes user churn.

Michele warns other founders against treating hands-off execution as the ultimate metric: “So I think autonomy shouldn't be fetish for agent builders at this point.”

- In September 2025, promoting 200 minutes of continuous autonomous execution revealed an unexpected flaw: users tuned out, lost context, and abandoned the app. - The March release pivoted to managed multi-agent swarms operating in parallel under the hood, speeding up task completion without confusing the user with raw complexity. - User churn is mitigated by maintaining active touchpoints during long runs, focusing human input on architectural planning, visual design, and scope definition.

---

[18:44](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1124s) Masking Latency with Education, Social Loops, and Gamification

Bianca Rangecroft details how Whering approaches the latency problem by fundamentally re-engineering the application experience while users wait for heavy background processing and query responses. When a user triggers an operation, the product avoids passive progress bars and instead routes them through three strategic design pillars: education, social interaction, and gamification. Rather than feeling stranded, the user is pulled deeper into the product ecosystem while the models work asynchronously.

```mermaid flowchart LR accTitle: Three-Pillar Wait-Time UX Framework accDescr: Three interactive UX strategies used to engage users during model processing.

Wait["Model Processing Latency"] --> Pillar1["1. Education"] Wait --> Pillar2["2. Social Interaction"] Wait --> Pillar3["3. Gamification"]

Pillar1 --> E1["Sustainability Mission"] Pillar1 --> E2["Feature Discovery"]

Pillar2 --> S1["Wardrobe Styling Challenges"] Pillar2 --> S2["Peer Advice ('Style My Suit')"]

Pillar3 --> G1["Duolingo-Style Micro-Tasks"] G1 --> Data["Feeds Data Pipeline"] ```

The educational pillar highlights the brand’s core sustainability mission and introduces alternative features within the digital wardrobe. The social pillar leverages community interactions, prompting users to jump into friends' closets to assemble outfits or respond to urgent peer style queries (such as styling a suit for an upcoming speaking engagement). Finally, borrowing proven engagement patterns from Duolingo, the gamified layer presents bite-sized, rewarding mini-challenges. These playful tasks not only make waiting feel seamless and invisible, but also systematically feed refined user preference data back into the underlying data pipeline.

Bianca captures the core goal of their engagement loops: “And then in terms of the gamification segment, again, we've sort of re-engineered the whole thing to take a trick out of Duolingo's playbook to think about, how can we get the user to complete really fun, engaging challenges that feel native and core to the experience so that they don't even remember that they're waiting.”

- Interactive wait states replace loading spinners with native, value-add micro-actions. - Social mechanics allow users to browse peer wardrobes and crowdsource outfit choices during processing intervals. - Duolingo-style gamified challenges simultaneously entertain the end user and generate structured ground-truth data to fine-tune backend recommendation pipelines.

[20:22](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1222s) The Spectrum of Autonomy: From Synchronous IDEs to 14-Day Missions

Following up on how users tolerate processing wait times, the discussion highlights ongoing A/B testing around interim engagement features designed to sustain long-term retention. When users are given productive, educational tasks in parallel while an agent runs, the final output almost feels secondary because key milestones have already been achieved during the wait.

This dynamic reflects a broader "spectrum of autonomy" emerging across AI development tools. On one end sits the standard synchronous coding workflow: developers open an IDE or terminal, issue a prompt, and wait a couple of minutes while actively monitoring the screen. In the middle lies fully asynchronous automation—such as trigger-based, on-demand pull requests that run unattended initially before requesting human review. At the extreme end of this autonomy spectrum are long-running workflows referred to as "missions." These completely discard synchronous interaction in favor of a consulting engagement model. A user might invest two to three hours upfront strictly scoping out requirements and receiving a cost estimate, then dispatch the agent to execute independently over days or weeks without human supervision. One cited mission ran continuously for 14 straight days, consuming nearly 150 billion cached tokens while migrating an active 50-million-line codebase from a legacy programming language to a modern one.

```mermaid flowchart LR accTitle: The Spectrum of Agent Autonomy accDescr: Three tiers of agent autonomy from synchronous IDE copilots to asynchronous PR bots and multi-day consulting missions. subgraph Sync["Synchronous (Minutes)"] A["Developer in IDE / Terminal"] -->|"Prompt & Wait"| B["Inline Code Assist"] end subgraph Async["Asynchronous (Hours)"] C["Automated Triggers"] -->|"PR on Demand"| D["Human-in-the-Loop Review"] end subgraph Missions["Long-Running Missions (Days/Weeks)"] E["2-3hr Upfront Scoping"] -->|"14-day autonomous run"| F["Massive Migration / 50M LOC"] end Sync --> Async --> Missions ```

- In synchronous setups, users stay fundamentally anchored to their workstation, waiting minutes between iterative turns. - In long-horizon missions, users act like clients of an external firm, requiring precise upfront scope and expected cost estimations before disconnecting. - As the speaker notes regarding tasks that stretch past a week: “And I think that a lot of companies are going to start building out these sorts of strategies around, well, if it's going to be 10 days, then we just need to completely rethink the interaction pattern entirely.”

---

[22:35](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1355s) Production Bottlenecks: The Search for Verification and Outcome Oracles

Reflecting on how multi-day missions mirror external consultancy work, Logan Kilpatrick observes that the industry faces a looming transition away from metered token billing toward outcome-based pricing. He then pivots the panel to the biggest roadblocks encountered when shipping autonomous agents into production today, asking whether existing bottlenecks must be solved by foundational model labs, application founders, or the broader ecosystem.

Michele Catasta breaks the ice by identifying automated verification as the single greatest bottleneck holding back agent autonomy. While agents possess the autonomy to plan and execute multi-day task paths, verifying non-trivial outcomes remains unsolved outside of highly disciplined software repositories. In an enterprise codebase with comprehensive automated test suites and mature CI/CD pipelines, an agent extending existing tests provides reasonable confidence; however, real-world legacy codebases rarely match that level of test hygiene. Furthermore, shifting to outcome-based pricing requires an objective arbiter. If the end customer is the sole judge of whether a task was completed satisfactorily, game theory dictates that users will claim dissatisfaction to secure refunds. Rather than relying on standard "LLM-as-a-judge" heuristics, the industry requires an absolute oracle capable of giving a definitive binary evaluation of an agent's finished work.

- Variable per-token pricing is inherently difficult to explain to business clients, making outcome-based contracts far more commercially appealing if completion can be reliably verified. - Relying exclusively on CI/CD pipelines and unit tests is insufficient for messy, real-world codebases that lack Google-level engineering maturity. - As Michele Catasta explains when outlining the missing ingredient for trusting autonomous agents: “I'm really talking about Oracle at the end of the work that the agent has done, that gives you the thumbs up or thumbs down.”

---

[25:30](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1530s) B2C Pressures: Scaling Precision, Fast-Moving Competition, and Design Bottlenecks

Representing consumer-facing and B2B2C products, Bianca Rangecroft outlines two distinct hurdles that prove surprisingly stubborn. The first is managing fierce competition in an era where technical barriers to entry have fallen. When her team launched, differentiation required curating proprietary datasets and training bespoke styling and personalization models from scratch. Today, frontier capabilities are widely commoditized and accessible out of the box, shifting the competitive battleground to fundraising, go-to-market speed, and acquiring shared target demographics without sacrificing output precision or brand equity.

The second bottleneck is an acute shortage of design bandwidth. With an active community of 10 million users, Rangecroft's team captures massive volumes of daily feedback, scraping Reddit forums and running social listening pipelines to map out exactly what features need to be built. While modern code-generation harnesses allow engineering teams to implement and ship backend features faster than ever before, product design and brand craft cannot be generated at the same velocity. Cultivating a distinctive visual brand that stands apart from an influx of copycats requires human design resources that are difficult to recruit and scale.

- The historical moat of building proprietary training datasets has been compressed by foundational models democratizing personalization capabilities. - Rapid engineering execution must be balanced against high precision, as a single wave of flawed consumer-facing outputs can degrade hard-won brand recognition. - As Bianca Rangecroft characterizes their scaling dilemma: “And we can build and ship things so much faster. But who's designing this stuff. When you're thinking about a really, really strong brand that can demarcate in the sea of new competitors.”

---

[27:23](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1643s) The Mindset Overhang and Democratizing High-Cost Video Agents

Addressing systemic enterprise adoption, the panel unpacks the concept of capability or "model overhang"—the lag between what frontier models can achieve and what existing software products actually leverage. While long-context coherence and autonomous self-validation still impose technical ceilings, current agent harnesses have matured significantly. In place of a technical capability overhang, the industry now confronts an "enablement overhang" or "mindset overhang." Global enterprises across regions like Poland, Japan, and Latin America—sectors historically perceived as slow technology adopters—are already autonomously generating and shipping nearly every pull request. In contrast, digital native companies often remain paralyzed in internal committees debating tool selections. Leadership teams risk an unpleasant shock if they fail to recognize that developer workflows fundamentally shifted months ago.

Closing out the discussion, Jay Wu highlights raw inference cost as the primary barrier preventing video generation agents from reaching widespread adoption. Leading video foundation models (such as ByteDance's SeaDance / Seaweed series) currently cost between $0.10 and $0.50 per second of generated footage. At that unit economics threshold, usage remains restricted to well-funded professionals producing mission-critical commercial assets. Reaching the early adopters and the broader consumer majority requires significant price compression to democratize autonomous video workflows for average creators.

- Organizational inertia, rather than tooling sophistication, determines which enterprises capture value from agentic development. - The competitive advantage in modern software engineering stems from leadership communication and change management rather than choosing between rival developer tools. - As the speaker cautions executives who are hesitating to implement autonomous agents: “The world changed probably five months ago. And now if you're like, basically not adopting this at full speed responsibly, then it's going to lead to I think, a oh, no moment in a couple months where you're like, wait, when did this happen.”

[30:29](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1829s) The Runaway Unit Economics of Agentic Video

```mermaid flowchart TD accTitle: The Agentic Video Retry and Cost Loop accDescr: Flowchart illustrating how generating one minute of video requires multiple critique loops and vendor retries, tripling compute assets and driving up production costs.

subgraph AgentLoop ["Agent Execution Loop"] A["Generate Video Segment"] --> B["Critique & Evaluation"] B -->|"Fails criteria"| C["Punishment & Retry"] C --> A end

subgraph CostImpact ["Customer & Business Outcome"] B -->|"Passes"| D["Final 1-Minute Video Delivered"] C -.->|"Consumes 3x vendor compute"| E["Soaring Cost of Goods Sold"] E --> F["End Customer Billed $30-$60"] F --> G["Buyer Hesitation & Churn Risk"] end ```

Building an agent that handles generative video introduces brutal economic realities because autonomous workflows require constant self-correction. Jay Wu explains that when creating an agent, you cannot simply make a single inference call and deliver the output. Instead, the agent has to perform retries, implement critique-in-the-loop steps, and apply punishment mechanisms when outputs do not meet requirements. Because of these repeated corrective cycles, producing just one minute of finished video for an end customer often requires spending approximately three minutes' worth of generation assets across underlying model vendors. This multiplier drives the company’s cost of goods sold up dramatically, creating a real pricing friction point with consumers.

The core dilemma is that even when the final output is good or satisfying, customers do not necessarily spend enough focused attention on the product to justify a premium price tag. Once a creator or small business sees an invoice of thirty or sixty dollars for a short social media clip, they immediately begin questioning whether they should continue purchasing it. Jay points out that managing this compute cost remains an ongoing, major challenge for startups operating in generative media. Responding to the unit economics squeeze, Logan Kilpatrick offers reassurance from the model provider side, saying, “I promise we're going to make better models and we'll make them cheaper over time for you to hopefully help solve this problem.”

- Delivering one minute of finalized customer video often burns three minutes worth of upstream vendor compute due to continuous critique loops and automated retries. - End-user costs of $30 to $60 per social video create immediate buyer hesitation, regardless of whether the output is acceptable or satisfying. - High vendor inference expenses keep the cost of goods sold (COGS) for media generation agents exceptionally steep.

[31:32](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=1892s) Moving Beyond Frame-by-Frame Video Understanding

Logan Kilpatrick pivots the discussion toward the broader evolution of autonomous agents, noting that agents are rapidly transitioning from processing purely textual data to understanding and seeing the world through native multimodality. Pointing out that models like Gemini were conceived from the outset as multimodal systems, Logan asks the panel whether a technological chasm has truly been crossed in generation or visual comprehension, and how those capabilities are actually altering product roadmaps.

Jay Wu responds that the current state of video comprehension inside agent systems still falls well short of what autonomous production requires. Today, most multimodal agents are effectively restricted to basic image understanding applied sequentially: they ingest and inspect video files frame by frame rather than evaluating continuous motion. Jay emphasizes that an AI agent does not look at footage the way a human video editor does on a timeline, observing frames flowing together continuously. That fundamental limitation leaves automated clipping and editing features feeling clumsy. As Jay explains: “So that makes the video editing capability on the agent much less than ideal.” Looking to the future, Jay notes that he is eagerly waiting for truly video-native comprehension models to emerge inside agent frameworks. Hearing this, Logan hints at what DeepMind is preparing behind the scenes, teasing that they have exciting video capabilities in the pipeline and promising to email Jay directly about what they are cooking up.

[33:29](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=2009s) Multimodal Coding Agents and Visual Abstractions

The conversation turns to how multimodality is fundamentally transforming coding agents and developer workflows. Users increasingly demand and expect that software engineering agents should not be confined to editing text files; they should be capable of programmatically manipulating images, editing video content, and interacting with specialized web media frameworks like Remotion. As coding agents evolve into general-purpose problem-solving assistants, developers are leaning into specialized skills that bridge programmatic generation with visual assets, allowing automated systems to directly author multimedia interfaces and video components alongside backend logic.

Beyond simply manipulating graphics, multimodality plays a crucial role in how developers comprehend complex system architectures. While development workflows have recently adopted Mermaid diagrams and Markdown for quick technical documentation, relying on basic diagrams and raw source code is becoming a bottleneck for human comprehension. Most engineers no longer need to spend their days scrutinizing every single line of code to understand software behavior. By operating at a significantly higher layer of abstraction, rich visual interfaces allow developers to grasp and orchestrate deeply intricate architectures much faster than traditional text buffers allow. As the speaker highlights: “And I think that if you stay at that higher level of abstraction, visuals are one of the ways that we can build an understanding of even more substantially complex systems.”

## Notable quotes > [02:10](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=130s) The agent roadmap moves at a crazy pace, and we throw away part of the software we architected on a regular basis because frontier models make leaps every three months or so. — Michele Catasta

> [15:47](https://www.youtube.com/watch?v=RA_fQvXQ4aw&t=947s) I think autonomy shouldn't be fetish for agent builders at this point. — Michele Catasta

## Mentioned Logan Kilpatrick, Michele Catasta, Bianca Rangecroft, Jay Wu, Google DeepMind, Google Cloud Next, Gemini, Veo, Replit, Whering, Factory, Droid, GPT-3.5, Duolingo, Marc Andreessen, Ruby on Rails, Django, ByteDance, SeaDance, Seaweed, Remotion, Mermaid, Reddit, Disney

## Video description In the AI world, your only real moat is iteration speed. Logan Kilpatrick and founders from the front lines will share their recipe for moving fast without breaking the model. We’ll look at real workflows using Google DeepMind’s models, the technical trade-offs of multimodality, and how to stay ahead of the next model release.

Watch more Google Cloud Next 2026 Sessions → https://goo.gle/gcloud-next-2026 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech

Speakers: Logan Kilpatrick, Michele Catasta, Bianca Rangecroft, Jay Wu

BRK2-157 #GoogleCloudNext

[Source](https://www.youtube.com/watch?v=RA_fQvXQ4aw)

Reading summary

Speaking at Google Cloud Next, startup founders and Google DeepMind's Logan Kilpatrick highlighted that deploying AI agents at scale requires shifting from thin API wrappers to deep architectural harnesses that manage long-running execution, human-in-the-loop engagement, and severe cost or evaluation bottlenecks. Architectural Tension Between Fast Agents and Deep Infrastructure (2:10) Building agentic software exposes a structural conflict between slow infrastructure and hyper-iterative frontier models. Michele Catasta explained that Replit spent eight years constructing developer sandboxes and security controls before adopting an agent-only architecture for non-technical knowledge workers. While core data infrastructure must evolve deliberately, the agent layer is frequently refactored or discarded every few months to absorb new frontier model capabilities. Similarly, Factory's speaker noted that their autonomous software development lifecycle 'droids' required rewriting tactical integrations and runtime abstractions for two and a half years before underlying model reasoning became adequate. Rethinking Latency and the Spectrum of Autonomy (15:47) Long-running autonomous execution creates critical user-experience challenges when users abandon interfaces during extended processing. Replit reported that a release featuring 200 minutes of continuous autonomous execution backfired, prompting a shift toward parallel swarms running in the background while users stay engaged in architectural planning and design. Whering addressed processing latency by embedding educational sustainability content, social wardrobe sharing, and Duolingo-style gamified challenges that simultaneously gather preference data. Across development tools, tasks now span a spectrum from synchronous IDE prompts and asynchronous pull-request reviews to multi-day 'missions' requiring upfront scoping, such as a 14-day continuous code migration consuming nearly 150 billion cached tokens. Verification Oracles, Brand Craft, and Organizational Mindset Overhang (22:35) Scaling production agents reveals distinct commercial and operational bottlenecks across sectors. In enterprise codebases, Catasta argued that shifting to outcome-based contracts requires objective 'oracles' capable of binary verification, because real-world repositories lack Google-grade test suites and clients face economic incentives to dispute subjective outcomes. In consumer e-commerce, Bianca Rangecroft noted that commoditized models compressed the moat of proprietary training data, moving the bottleneck to scarce human design talent needed to maintain brand distinction. Enterprise adoption is further constrained by what panelists termed an 'enablement' or 'mindset overhang,' where organizational inertia and committee debates stall deployments even as non-traditional tech firms abroad automate pull requests. The Unit Economics and Comprehension Limits of Multimodal Video (30:29) Creative video agents face steep economic and technical barriers compared to text-based tools. Jay Wu explained that raw video foundation models default to generic outputs, necessitating curated storytelling templates and costly human-rubric evaluations to instill aesthetic 'taste.' Moreover, autonomous self-correction mechanisms—including automated critique, retries, and punishment loops—mean delivering one minute of finished video often consumes three minutes of upstream vendor compute. With foundation model costs ranging between $0.10 and $0.50 per second of generated footage, end users face invoices of $30 to $60 for short clips, depressing purchase retention. Video agents are also limited technically by evaluating footage frame by frame rather than comprehending continuous motion on a timeline. While coding agents increasingly serve as general-purpose computer harnesses that manipulate media alongside logic, unlocking widespread consumer and enterprise adoption depends on lowering model inference costs, developing native video comprehension, and building reliable evaluation oracles.

Source quotes

[2:10] Michele Catasta: The agent roadmap moves at a crazy pace, and we throw away part of the software we architected on a regular basis because frontier models make leaps every three months or so. [15:47] Michele Catasta: I think autonomy shouldn't be fetish for agent builders at this point.

Original transcript

[0:13] How's it going everyone. [0:14] I'm excited for this panel. [0:16] I'm excited to have a wonderful group of folks representing [0:19] four different companies, pushing [0:20] the frontier in different ways. [0:22] Hopefully everybody's next has been fun. [0:24] Hopefully you've got to listen to lots of things, [0:27] hopefully meet a bunch of cool people, [0:30] maybe gamble a little bit. [0:31] I don't know what people do in Vegas besides gamble [0:34] and commit to these talks, so hopefully it's been fun. [0:38] And yeah, my name is Logan Kilpatrick. [0:41] I lead the developer platform team in DeepMind [0:43] focused on building our APIs, bringing the models to you [0:46] all so that you can build on top of Gemini app and Veo [0:50] and the rest of our suite of models. [0:52] So it's a pleasure to be here. [0:54] It's a pleasure to have this conversation. [0:57] And I will be asking tons of random questions [1:00] really, about how four companies that [1:03] are on the frontier are using are using our technology, using [1:06] the technology in general. [1:08] And hopefully we'll get a bunch of Super interesting information [1:11] out of you all. [1:12] And then we'd love to do some Q&A. So if folks have questions [1:16] for our panelists would love to. [1:18] Yeah, we'd love to hear what you all have to. [1:21] Yeah, we'd love to hear what you all have questions about. [1:25] Before I do that, let's do I think we see everybody's names [1:29] up on the screen actually. [1:30] So maybe we'll skip intros and we'll dive in to questions. [1:34] I feel like the thing that's top of mind, [1:37] I think, for lots of folks right now, [1:38] is obviously all the things that are happening in agents [1:41] and coding. [1:42] And we have two coding products and we [1:44] have two products that have agentic capabilities built [1:48] on top of them. [1:49] And so I'm curious how you're thinking about this problem [1:54] space and also how that's changed like from a product [1:57] perspective, actually in the implementation [1:59] details over the last six months, [2:01] as everyone has sort of evolved their product [2:04] to become like so much more. [2:05] Agentic and Michaela with rap-a-lot maybe. [2:07] Do you want to start. [2:10] Yeah, we made that bet like a year and a half [2:12] ago that our product was exclusively an agent, [2:15] and it turns out that it was the right bet to make. [2:18] Can you give context why was that the bet. [2:20] Why was it so clear a year and a half ago. [2:24] That's the direction you wanted to go. [2:26] I think it was mostly conviction and not clarity [2:29] back in those days, because agents were not nearly as [2:31] powerful as they are today. [2:33] So the outcome you were seeing was [2:36] exciting for someone who understands [2:37] research and exponential, maybe not as [2:39] exciting for people who actually need to ship with agents. [2:43] But we knew that would have happened really fast. [2:45] Maybe a few months after the first launch [2:48] we made, it was a fundamental choice for us [2:50] because we target knowledge workers. [2:52] We don't target software developers, [2:53] so we don't want to put them in front of an idea. [2:56] An idea is the wrong product service for someone [2:58] who's not non-technical. [3:00] And since then, I think the product philosophy [3:02] hasn't changed. [3:03] What changes how ambitious we get in terms [3:06] of what the agent should build. [3:08] And we started with relatively simple prototypes. [3:11] And now we go to Enterprise customers. [3:13] And they create internal tools used by thousands [3:15] of employees on a daily basis. [3:17] And those tools are scaling well. [3:18] They have data governance and security requirements [3:21] that make them happy. [3:23] So the entire field is moving this fast fundamentally [3:27] because AI has become more powerful. [3:29] And even with that early bet, how much like has it [3:32] been a huge amount of effort and engineering energy [3:36] to rebase a bunch of the stuff or the infrastructure [3:39] that you invested in a year and a half, [3:41] has actually been the infrastructure that's [3:42] carried over and you've just been able I'm simplifying, [3:46] but just add more tools to what the agent can do. [3:49] And it's more robust now for example. [3:52] So it is a 10 year overnight success, [3:55] which means eight years were invested [3:57] on building infrastructure and the tools the developers use. [4:01] And suddenly we were at the point [4:03] where we could build an agent that [4:05] turned to be the connective tissue among all the tools [4:08] that we had in the past. [4:10] And we built it on top of it. [4:11] So we had a massive advantage by having [4:14] spent a lot of time building infrastructure. [4:17] The agent roadmap moves at a crazy pace, [4:21] and we throw away part of the software we architected [4:24] on a regular basis because frontier models make leaps [4:27] every three months or so. [4:29] At this point, the infrastructure roadmap [4:31] moves at a totally different timeline. [4:33] There are certain things that need [4:35] to move with a more sustainable pace, because you [4:38] need to keep security in mind. [4:39] You need to take care of your user data, [4:41] and it creates a very interesting tension [4:43] within the company. [4:44] When you have these roadmaps that [4:46] are not moving at the same pace as you, I'm sure you know. [4:48] Yeah, yeah, lots of tension in that direction. [4:51] Jay, I'm curious, actually, on the multimodal side, how you all [4:55] are thinking about this. [4:56] As somebody who's the product, bring a bunch of media, [5:00] generate a bunch of assets from a bunch of input media. [5:03] I'm curious, has this agentic era [5:06] been sort of a rebase for you, or was it literally this [5:09] was what you were betting on originally. [5:11] And it's just a different implementation detail [5:13] of how you all are thinking about this. [5:16] Yeah the agentic would bring much more capability [5:21] of understanding what the user want. [5:24] Because before that, the only thing you can do [5:28] is you wrap around the API and then [5:30] you let the user either you ingest their intention [5:35] and then wrap a prompt around them [5:38] and then prompt the model to give them certain multimedia [5:42] assets. [5:43] Nowadays you can understand their high level goals. [5:47] You can use to do research and build a full story [5:51] and then create a video for them end to end. [5:54] And you can also take feedback from them [5:56] to do human in the loop iterations, [5:59] so that it's much more exciting era [6:02] where you can deliver so much more value to the users [6:05] than reselling APIs in some sense. [6:09] Yeah, no, that makes a ton of sense. [6:11] And Bianca, I'm actually also curious for you [6:13] all like this agent era, how much has this changed. [6:17] I think you all have a super interesting product [6:19] and strategy. [6:20] And so I'm curious how much this has [6:21] changed how you all are thinking about the product strategy. [6:25] Yeah, I mean, honestly, when we first [6:27] started pitching this in 2020, like an agent that [6:30] could be a personal stylist in your closet, a lot of investors [6:34] were like, this is never going to happen. [6:36] So we've come a long way. [6:38] I think for us, we always thought [6:40] that this time would come, which is a great thing, [6:42] that it actually panned out. [6:44] And we prepared the strategy in three different ways. [6:47] And I think it actually has really worked in our favor. [6:50] The first was thinking about how do we build this huge data set [6:54] in terms of what people are actually wearing, [6:57] what brands they have in their wardrobes, what are their sizing [7:00] and styling preferences so that flywheel, [7:03] it was intrinsic to all of the product led growth [7:06] around the digital wardrobe app, and that [7:08] was really great in terms of building our moat, [7:11] trying to then think about how we can build really [7:13] niche specific models that can hit [7:15] a lot of these different and new deliverables [7:17] that the customers actually want, [7:19] and they're getting more and more savvy in terms [7:21] of what they want, US as a company [7:23] to do for them, whether it's packing for trips, planning, [7:26] optimizing your wardrobe, giving insights [7:28] to potentially verify new purchases, all of that stuff [7:32] has kind of been brought into the thinking around the outputs. [7:36] The second thing we've always been thinking about [7:38] is basically different capabilities, [7:40] how can people try on clothes, right. [7:43] So we've been thinking a lot about what [7:45] is the data that we need to amass as a company [7:47] to make virtual try-on really, really [7:49] detailed and personalized. [7:51] And that's been stuff like getting users [7:53] to log pictures of what selfies of what [7:55] they're wearing on a daily basis. [7:57] We've built out a huge understanding of the consumer's [8:00] body in different clothes, in different moments, [8:03] and that's really been able to help us fine tune, I guess, [8:06] the product outcomes that we're looking for. [8:09] And then thirdly, our big bet was always [8:11] to say the B2C flywheel is one thing. [8:13] And we're working on, incredible organic growth there, [8:16] going into new markets like the US LATAM. [8:19] But really the big vision was always B2B. [8:21] How can we leverage this data set in order to personalize [8:25] the e-commerce experience. [8:26] And so I guess now something that's [8:29] incredibly top of mind for retailers. [8:31] And it comes at a really great time for us [8:33] as we've just closed a seed round. [8:35] That's awesome. [8:36] Well congrats. [8:38] I'm curious for you all at factory building droids, which [8:42] I love. [8:43] I love the droids thing, not agents. [8:45] They're droids. [8:46] Actually, I'm curious internally, are they agents [8:48] or are they droids? [8:49] Droids are like the. [8:51] No, no, we use the term droids because we started [8:54] three years ago when GPT 3.5 came out and we said, [8:58] we want to build fully autonomous AI systems [9:01] for enterprise businesses that work across the entire software [9:05] development lifecycle. [9:06] And that means that we have to build [9:08] a layer of governance and management [9:10] of these autonomous AI systems. [9:13] And it just didn't work for 2 and 1/2 years. [9:15] And so I think that in terms of seeing what was going to happen, [9:22] we definitely had a lot of iteration time [9:25] to basically imagine it. [9:27] But it's very different to imagine what that would [9:30] look like at a very high level. [9:32] I sometimes joke, you could literally just [9:34] take our seed pitch deck and present it in a meeting today, [9:38] and the people would just nod and be like, great, sounds [9:40] great. [9:40] Let's get it deployed because it just basically [9:42] is the same vision. [9:44] But the tactics of how we got there [9:46] were so varied along the way. [9:48] The world just changed many times. [9:50] Model capabilities came out. [9:51] How you actually integrated that in, [9:53] it just kept changing and changing and changing [9:55] and so Droid really came out of the fact [9:58] that the term agent didn't exist yet. [10:00] I mean, it did, but only in cognitive science literature. [10:03] So we didn't think that was going to be something [10:05] that necessarily caught on. [10:06] And so we said they're droids. [10:08] You had a different name. [10:09] And then Disney stopped you from using that name. [10:12] Allegedly No, no, no. [10:16] The name has always been Droid. [10:18] I think that you're actually see ask me [10:20] that question again in a year and then [10:21] you might get an answer. [10:23] So far lawyers in the room or we are outside of merchandising. [10:28] So I think it's probably fair game to keep it for now. [10:30] I love it. [10:31] I'm curious and actually for you all to plant this seed, [10:34] I'm curious as the agents start running longer. [10:38] I'm curious how this will impact all your products [10:41] and businesses. [10:42] But actually before that, I'm also [10:44] curious as coding becomes this general purpose agent harness [10:49] potentially and it's like code all the way down [10:52] and code generated for all these use cases. [10:55] I'm curious how you've has that evolved the way that you all [10:58] have thought about what the actual product strategy is. [11:00] It's like less in some cases potentially [11:02] about building software. [11:03] It's like, could just be the Droid [11:06] could just be a general purpose agent that any other company is [11:08] using, even if they're not doing software development tasks. [11:11] Is that something that you all are thinking about or excited [11:13] about. [11:14] Yeah, totally. [11:15] That's like the million dollar question. [11:17] Honestly, that's the billion dollar question [11:20] that I think is interesting because what we've seen [11:23] is that Droid is so competent at effectively navigating [11:26] the digital world because underneath [11:28] of everything we do on a computer is software. [11:31] And I think that the innovations in self recursive improvement [11:34] on coding agent harnesses have pushed them [11:37] to the point where if you do have an extremely [11:39] competent, local, first model agnostic coding agent, [11:44] you kind have a extremely competent general agent [11:47] for almost any work that can be done on a computer, [11:50] and that's most work. [11:52] And so we've thought a lot about that. [11:54] And I think that right now we're laser focused on making sure [11:56] that enterprise software organizations are [11:59] able to across their entire software business, [12:02] do the things that they need to do. [12:03] And sometimes that means product managers, data scientists, [12:06] product marketers. [12:07] Our third highest user internally of Android [12:10] is an account executive on our team selling Droid with data. [12:14] And I think that means that there's a lot of opportunities, [12:17] but we don't see it as us moving into other spaces. [12:21] Instead, we see it as other spaces becoming software. [12:23] And I think that that's something [12:25] that we've known for a long time, [12:26] is that software is eating the world. [12:28] And so a lot of what we do is say, let's work on software. [12:31] And it just so happens that software is kind of everything [12:34] is computer, so to speak. [12:36] Yeah, I love that. [12:38] We could talk about this thread for a long time, Jay. [12:40] I planted the seed about of long running agents. [12:43] We were talking before we got on stage about the tension for you, [12:47] all of this taste human element in your product experience [12:52] and how it definitely still requires [12:53] in certain cases like this human in the loop [12:55] to make a judgment decision. [12:57] And I'm curious as agents out of the box start [13:00] going longer running operations. [13:02] Et cetera. [13:03] Et cetera. [13:05] What is it still going to be taste. [13:06] That's the bottleneck for you all. [13:07] Or, where do you think the bottleneck [13:08] is going to start to move in the product experience [13:11] that you all are building as agents start running longer. [13:14] So first of all, we are really excited [13:16] about the development of harness, [13:20] because before that you have to write, [13:23] basically write your own agent React loop and deal [13:26] with all these things and context engineering. [13:29] Now are better and better harness. [13:32] It fuels me to that. [13:34] It feels like the MVC framework during the web era. [13:38] Before that have to write all these things by hand, [13:40] and then you have Ruby on Rails and Django, [13:43] and it's much easier to write reliable software [13:46] and to serve the customers. [13:48] So with that have to focus on the taste [13:52] because we serve the customers by helping [13:55] them create good social media ready videos as agents. [13:59] However, social media is a very competitive arena, [14:03] and you cannot just make a slob and hope to get a lot [14:07] of viewership. [14:08] You have to tell good stories. [14:10] You have to tell, give good visuals, [14:13] and all these have to stand on top of what everybody else has [14:17] been doing on social media, which [14:19] means that the foundational model has [14:21] their generic and homogeneous taste [14:24] if you just use them without a lot of thoughts into it. [14:29] And one of our job is to inject a lot of the taste [14:33] on top of in our product, on top of foundational multi-modal [14:39] generators. [14:40] And it's very important to measure the quality [14:46] of the output of your product. [14:49] So when you're iterating these skills and iterating templates, [14:54] you have to make sure that any iteration brings [14:58] your product better, not worse. [15:01] And without a good eval mechanism, [15:04] it's a lot of guesswork. [15:07] And once you ship something, you don't [15:09] know whether some people will get worse results, [15:11] and maybe some people will get better. [15:13] And it has always been a challenge in this [15:17] because it's easier to detect defects in videos, [15:20] but it's hard to measure taste. [15:23] And you have to combine human and a lot of rubrics, rubrics, [15:27] engineering to measure these to measure [15:30] the quality of the output. [15:32] And they can be expensive and inaccurate. [15:35] So these are the challenges still being [15:39] quite being there during agentic era for video Gen agents. [15:45] Yeah no that makes a ton of sense. [15:47] We were also talking before we got [15:49] on stage sort of about this speed, [15:52] to accuracy tension that exists. [15:54] And I'm curious for the other three panelists, [15:56] like on this same through line as agents [15:58] start running longer and longer and longer and doing more, [16:03] it's obviously going to take longer to make that happen. [16:06] Unless all of a sudden we have magic chips that [16:08] make everything super fast and the tension is like, what. [16:12] What does the user do while this thing is out running and doing [16:15] a bunch of work on your behalf. [16:16] And I'm curious if anyone has thoughts on the product [16:21] experience, innovation or the way [16:22] that you're breaking apart problems [16:24] to make it so that even if the model is capable of going [16:27] and doing 24 hours of work, like the user [16:30] wants to see something and know that there's things happening [16:33] and that the tension between the long running tasks [16:36] and the immediacy of everyone. [16:38] And I'll share my personal experiences [16:39] like we've seen this across our products, which [16:41] is the faster you give users an answer, [16:44] it's the whole original Google Search thesis. [16:46] If you can do something in 100 milliseconds, people are happy [16:49] and they stick in the product and they love it. [16:51] If you make them wait tens they drop off [16:53] and they go do something else. [16:54] And I feel like as we move to even longer, [16:56] there's a lot of tension in there. [16:58] And I'm curious how y'all are thinking about this. [17:01] We disrupted our entire product experience [17:04] when we launched three back in September 2025. [17:08] We were marketing the fact that our agent [17:09] was running for 200 minutes in a row autonomously. [17:12] I think we were the first company [17:13] to just start to talk about the fact [17:15] that it could run for hours, and it turns out [17:18] that we probably didn't put enough emphasis on the fact [17:21] that in reality, users will start to not pay attention [17:24] to the product anymore for that reason. [17:26] So I think autonomy shouldn't be fetish for agent builders [17:31] at this point. [17:32] Yes, it's important that you take a user from point A [17:34] to point B without having them to worry about any [17:37] of the technical details, but at the same time, [17:40] the product experience has to be such that you remain engaged. [17:43] And I think if purely you focus on making the agent more [17:47] autonomous and there is nothing else [17:48] to do in your product service, you're [17:50] going to lose them in the process. [17:53] What we did with the latest launch back in March this year [17:56] is we allow you to do a lot of planning in parallel where [17:59] several agents are running. [18:01] Basically, we have a swarm that is completely managed by us, [18:04] and our users don't even understand that [18:06] that's happening, most likely. [18:07] So there is a way to make things go faster [18:09] by means of parallelization, and there [18:11] is a way to keep users engaged by just giving them more reasons [18:15] to work on your product, which could be planning in our case [18:18] is also a lot of designing happening. [18:20] A lot of other things that are about the software development [18:23] lifecycle. [18:24] And I believe every product out there, [18:26] even if they're not coding agents, [18:28] they should start to think in these terms [18:30] where yes, you will have agents running in parallel. [18:32] They are going to be running autonomously, [18:34] but you need to make sure that the user is [18:36] engaged in some other ways. [18:38] Otherwise, exactly what you were mentioning, [18:40] they're going to drop. [18:41] Yeah and the drop is savage. [18:44] Yeah So I think from our perspective, [18:47] we're thinking about re-engineering the user [18:49] experience in the app in three different ways [18:53] as they wait for the feedback that they're expecting. [18:56] The first is thinking about education, [18:58] the second is social for us, and the third is very obvious. [19:02] But gamification. [19:04] And so in terms of the education piece, [19:06] everything that we're trying to build now [19:08] is you wait for a response to a query, [19:10] or you're expecting an output from the app [19:12] is asking you to either give it more information [19:14] or data in a really fun, gamified way, [19:17] or is educating the end user on why this is important. [19:21] What else you can do with the app in the meantime. [19:23] Thinking a lot about sharing our mission, our vision, which is [19:26] centered around sustainability. [19:27] And so we found really fun and engaging ways [19:29] of doing that and hitting product education [19:32] while the user waits. [19:34] And then when we think about social again, for us, [19:37] we're a social wardrobe app. [19:38] So can go into your friends wardrobes, [19:40] style them, you can post style challenges with I don't know how [19:43] to wear the suit, guys help. [19:45] I got a big talk tomorrow on stage. [19:47] And so we think about leveraging those challenges [19:52] as the user waits in order to gamify that experience [19:55] and get them doing fun, reinforcing things that continue [19:59] to feed that data pipeline. [20:01] And then in terms of the gamification segment, [20:03] again, we've sort of re-engineered the whole thing [20:05] to take a play, sorry, a trick out of Duolingo's playbook [20:09] to think about, how can we get the user to complete really [20:13] fun, engaging challenges that feel native [20:16] and core to the experience so that they don't even [20:18] remember that they're waiting. [20:19] But they actually get brought into something else. [20:22] And we're a B testing right now. [20:24] What are the best things that actually drive long term [20:26] engagement and retention. [20:27] And so we find that we can have these two things existing [20:30] in parallel. [20:31] And then they're really excited to get the final output [20:33] when it's ready. [20:34] And it's almost secondary because they've hit these three [20:37] things in the interim. [20:38] I love that. [20:39] Yeah, first of all that's great. [20:41] I think the education work in progress work in progress. [20:44] No but that's the education on the wait [20:46] times is like a fantastic idea. [20:48] I think that we think of there being [20:51] like the spectrum of autonomy, where a lot of people [20:54] are using synchronous coding agents, [20:57] almost like in their daily workflow. [20:58] They open up a terminal or an Ide. [21:00] They say, hey, could you please do this thing. [21:02] And the expectation is maybe some degree couple minutes [21:05] and they're managing, or even if they flip to another screen, [21:09] they're still fundamentally there. [21:11] And then you do have this world of fully asynchronous, [21:15] pull requests on demand, trigger based automations [21:19] that don't even need a human in the loop [21:21] for the initial component, but may then [21:22] need some sort of engagement thereafter. [21:25] But on the very far end of the spectrum, [21:28] we have a product called missions, which is effectively [21:30] completely breaks the synchronous paradigm by enabling [21:35] a user to define a scope of work that can be quite substantial. [21:38] The longest mission ran for 14 days [21:41] straight of continuous work, used almost $150 billion cash [21:45] tokens. [21:46] And ultimately, workflow is more reminiscent of working [21:50] with a consulting firm almost, where you say, hey, look, [21:53] I'm going to spend all of my synchronous time [21:55] up front for maybe even two or three hours figuring out [21:58] precisely what I want, the scope of work to be done. [22:01] And I almost expect a cost estimate [22:03] of what that's going to look like. [22:04] And then I send off this work to be executed, [22:07] and I know that I'm going to be not present. [22:10] And this was almost 50 million line [22:12] code base where there was a migration that [22:14] was going actively across that entire code base [22:17] from a legacy language to a more contemporary one. [22:20] And I think that scale of work, we [22:21] expect to be very powerful and different. [22:25] And I think that a lot of companies [22:26] are going to start building out these sorts of strategies [22:29] around, well, if it's going to be 10 days, [22:31] then we just need to completely rethink the interaction pattern [22:34] entirely. [22:35] Yeah that's so interesting. [22:36] I think there's been lots of conversations in the last two [22:39] years about actually to the point of this consulting model [22:43] and how the pricing will change. [22:45] Do you really want to be charging [22:47] for tokens in that context. [22:48] Is it an outcome based pricing, et cetera. [22:51] It's going to be interesting to see how [22:53] that to see how that plays out. [22:55] I have another question, which is [22:56] I'm curious as we sit here last year, there was lots of problems [23:01] as far as putting AI in agents into production. [23:04] A lot of those things hopefully have gotten easier [23:06] for you all and for other folks who are building stuff. [23:09] I'm curious where the bottlenecks are today across [23:13] all four of your different products [23:14] and how you're thinking about is it [23:19] are these bottlenecks problems that you all are taking on. [23:21] Is it things that you're like the ecosystem is [23:23] going to solve this and you'll benefit from those things. [23:26] Do you need the model companies to solve them. [23:28] I'm curious, what are the top of mind bottlenecks. [23:33] I don't know if anyone has anything top of mind, [23:35] or maybe life is good and there's no bottlenecks. [23:39] Plenty of problems. [23:40] I can break the ice. [23:42] I think the biggest bottleneck is verification, and the idea [23:45] that agents at this point are autonomous enough [23:48] to create paths to work for days in a row. [23:51] But fundamentally, if we knew how [23:52] to verify the outcome of what an agent has done outside [23:55] of the realm of coding, by the way, because if you work [23:57] in an enterprise code base and it's very thoroughly [24:01] tested and your agent is just extending the test suite, [24:04] you probably get some level of certainty [24:07] that your code is functioning correctly. [24:09] But we both know how it works in the real world, [24:12] and especially for code bases that are not [24:13] as mature as Google, then it's hard to rely exclusively [24:17] on tests is hard to rely exclusively on Ci, CD. [24:20] So if we had proper verification techniques both with [24:24] and without coding in the picture, [24:26] it would really help us to first of all, trust agents running [24:30] autonomously for long. [24:31] And at the same time also change our pricing model to your point, [24:35] we don't want to charge per tokens anymore. [24:36] It's hard to explain. [24:38] It's a variable cost. [24:40] It would be much better to charge by outcome. [24:43] But how can you verify the outcome as being accomplished [24:46] or not if the user is a judge of the outcome, then, of course, [24:51] in terms of as a game theoretical framework, [24:54] they will say, no, you didn't do what I wanted. [24:56] And I want a refund. [24:57] So I think it's a big AI problem. [24:59] I know most of the labs are working on that. [25:02] It's hard to see exactly what's going to be [25:04] the shape of a proper verifier. [25:06] I'm not talking about LLM as a judge. [25:08] That's a technique that has been around for a while. [25:10] Of course, it helps us to make judges more powerful. [25:12] I'm really talking about Oracle at the end of the work [25:15] that the agent has done, that gives you the thumbs up [25:17] or thumbs down. [25:18] So if you're working on the Holy Grail, please let us [25:20] know you want AGI is what it sounds like. [25:23] We all want AGI on this stage. [25:25] Yeah, you'll get it. [25:26] You'll get it. [25:27] Yeah I'm curious. [25:28] Other bottlenecks. [25:29] That makes sense. [25:30] Yeah so I think for us, being kind of predominantly B2C or B2B [25:35] B2C, sorry, there are two things that are weirdly, really [25:39] difficult to fix in a world where everything is fixable. [25:43] For us, the first is, I guess, competition generally. [25:46] When we started this, it was all about building the data set. [25:49] It was all about trying to train your own models [25:51] to do some form of styling or some form of personalization. [25:56] And today, obviously a lot of this stuff [25:57] is kind of open to all. [25:59] And so keeping up with the pace of competition in a B2C world [26:03] at cost, when you're looking at who is raising more money [26:06] and therefore has more to do, proper [26:08] go to market to think a lot more about acquisition of potentially [26:13] a shared target market. [26:15] So that's been really interesting in terms [26:16] of thinking about how we can ship at scale really, [26:19] really fast with precision. [26:22] And I think that's the key point. [26:23] Because this is a whole new world where [26:25] if you're thinking about your brand awareness, your brand [26:28] presence and recognition, ultimately your outputs [26:31] have to be really good. [26:32] And so that's been really interesting to adapt [26:35] to think about how do we build faster and ship faster [26:38] with all of the new models coming out [26:39] and all of the immense capabilities [26:41] that we've been waiting literally years for. [26:44] And then the second thing, which is interesting, [26:46] is thinking about design resource. [26:49] So we have a lot of feedback. [26:52] A community of 10 million users. [26:54] We get literally iterations or suggestions coming through [26:58] on a daily basis. [27:00] We can scrape, forums like Reddit [27:01] do a lot of social listening. [27:03] We kind of know where the product capabilities [27:05] need to go and evolve to. [27:07] And we can build and ship things so much faster. [27:09] But who's designing this stuff. [27:10] When you're thinking about a really, really strong brand [27:13] that can demarcate in the sea of new competitors. [27:16] So that's something that we're trying to wrap our head around [27:19] in terms of staffing and hiring. [27:23] Yeah, I think that the thing that I would point to the most [27:27] would be so there's this concept of the model overhang [27:30] or capability overhang, where the models are just much further [27:34] ahead than the products that are able to take advantage [27:37] of the full capabilities of the model. [27:40] I actually think that that's still true, [27:42] but it's much less true today. [27:44] I think that we're clearly hitting some limits that [27:47] are related to coherence over very long periods of time [27:52] around the quality and performance. [27:53] About self-validation. [27:55] And I think what that means is that our harnesses are actually [27:58] the frontier harnesses that agent labs are putting out [28:01] are just starting to get quite good. [28:03] And now I think what we're seeing [28:04] is a mindset overhang or an enablement overhang, [28:07] where there are organizations in the Fortune 500 [28:11] that have just fundamentally different software development [28:15] organizations as a result of the deployment of these systems. [28:18] And it's not really geographically split [28:21] like in my head. [28:22] I'm thinking about some of our customers in Poland, in Japan, [28:26] in LATAM, there are organizations [28:29] that are traditionally would probably [28:31] say, quite slow on tech that are almost fully autonomously [28:36] shipping almost every single one of their PRS. [28:38] And then there are organizations that are digital natives today [28:42] that are still debating should we use this tool [28:44] or should we use that tool. [28:45] And so to me, this seems like it's actually a mindset thing. [28:48] There are organizations that just haven't fully [28:51] grokked how the world is today, not how it will be in six months [28:56] or how everything's going to change soon. [28:58] The world changed probably five months ago. [29:00] And now if you're like, basically not adopting this [29:04] at full speed responsibly, then it's going to lead to I think, [29:08] a oh, no moment in a couple months where you're like, [29:11] wait, when did this happen. [29:12] When did everybody adopt these tools. [29:14] And so I think that mindset is incredibly important for a CIO, [29:19] CTO leadership team to have. [29:22] And we do need to go full steam. [29:24] The tools probably matter, at least to the extent [29:26] that you have a frontier harness. [29:28] But beyond that, all of the alpha [29:30] is going to be how you choose to communicate to your team [29:33] how to use the tool, not whether it's like tool a or tool B. [29:38] Yeah, for us, I think the cost is still a blocker. [29:43] It has been quite high and so far, [29:46] these best IAM models on videos are ranging from $0.10 to $0.50 [29:52] per second, which means that mostly the professionals can [29:56] afford it. [29:57] So you can see that CC dance has been C, [30:00] dance two has been pretty famous and people are using it [30:04] and they are used for mission critical work at the moment. [30:08] So if we want to have more as we tap early adopters [30:15] as well as the early majority, which is about democratized way, [30:21] democratize video agent. [30:23] You want to make it more affordable for average person [30:27] to be able to use it. [30:29] And when you are building an agent, [30:31] you inevitably have to do retries. [30:34] You do critique in the loop, and then you do punishment [30:38] so that every minute of video you deliver to the customer, [30:41] you probably need to spend three minutes worth of assets [30:45] spend on vendors, and that makes your cost of goods [30:50] sold incredibly. [30:51] Incredibly high. [30:53] And even at good quality people may not feel it's worth it. [30:58] They are not spending too much attention on your product. [31:01] They get a OK video or. [31:04] A satisfying video. [31:06] But then when they see the bill and that [31:08] is $30 per video, $60 per video. [31:12] For their social media, they have a question mark [31:15] on whether they should continue to buy it or not. [31:19] So cost has been still a challenge. [31:23] I promise we're going to make better models [31:24] and we'll make them cheaper over time [31:26] for you to hopefully help solve this problem. [31:29] US too please. [31:30] Yeah, yeah. [31:30] Of course. [31:32] I've got one more question. [31:33] I've got many questions, but I'll do one more. [31:35] And then I would love to hear questions from the audience. [31:38] I think it's been as a thread of this conversation [31:43] has just been like agents doing more stuff. [31:45] And I think obviously agents are starting to not just understand [31:50] text and process text, but also see things [31:53] in this whole multimodal space that we're in. [31:55] Obviously for some products with styling and video understanding [32:00] for you all j as well. [32:02] I'm curious how you all are thinking [32:04] about this multimodal world and how much it's a part of. [32:09] Yeah, the way that you're thinking about it and actually [32:11] how much it's changed because for some at least for the Gemini [32:14] story, we've been multimodal to begin with. [32:16] It's been something that's been a focus. [32:18] But I'm curious if there's been like a chasm that's [32:20] been crossed for any of those capabilities, generation [32:23] or understanding and whether or it's [32:26] making an impact in any of the products. [32:30] Whoever wants it. [32:32] Well, I'm really looking forward for video understanding [32:37] capability inside the agent right now. [32:40] Most of the agent can do image understanding, [32:45] but that is so much it. [32:47] It's helping customers clip a video or making a video, [32:53] but you are looking at the film frame by frame. [32:56] You are not looking at a La editor [33:00] on the screen playing frames continuously. [33:03] So that makes the video editing capability on the agent [33:09] much less than ideal. [33:12] So I'm kind of looking forward to see more video native [33:17] capability from an agent perspective. [33:20] We've got something good coming. [33:21] I'll email you, Jay, some video stuff that we're cooking on. [33:26] So it's exciting. [33:27] It's cool. [33:28] Other folks. [33:29] Yeah, I think multimodal definitely matters. [33:32] I think a lot about some of the work [33:34] that we've started to do where you I think that users almost [33:40] expect that you can manipulate images, edit videos, work, [33:44] like the remotion. [33:46] And I think there's something like Hyper, [33:48] something hyper, some skills that are coming out [33:51] that are oriented around coding agents manipulating and creating [33:54] on that string of coding agents being [33:57] general agents like video content, image content. [34:01] And a lot of the world sort of operates. [34:04] We shifted to Mermaid diagrams and markdown. [34:08] But I think that if we actually had richer ways [34:10] to visualize complex systems, we would do a lot of us [34:14] don't need to read individual lines of code to the degree [34:18] that you used to I would say probably ever. [34:20] But, that's not true for everybody yet. [34:22] And I think that if you stay at that higher [34:24] level of abstraction, visuals are one of the ways [34:28] that we can build an understanding of even more [34:31] substantially complex systems. [34:33] And so definitely feels like it's important for our work.