Factory Deploys Claude Opus to Cut Enterprise Coding Work by 90%

Claude

Reliable long-horizon planning by frontier models eliminates developer micromanagement, forcing software teams to continually scrap custom agent scaffolding as base capabilities advance.

AI-assisted summary of How Factory is building autonomous software with Claude — original source. Check the original for full context.

Source published: 2026-10-23

Key ideas

Factory leverages Claude Opus to conduct long-horizon architecture and planning for autonomous coding agents ('Droids'), enabling enterprise software developers to trust generated plans with minimal oversight. (5:29)
Deploying Droids into a major bank for enterprise COBOL modernization reduced required human involvement by 90% after the initial onboarding pass. (4:09)
Factory conceptualizes software autonomy as three concentric circles—frontier model, agent harness, and enterprise workflow—warning that builders must be ready to tear down orchestration layers monthly as models continuously absorb scaffolding functions. (8:12)
Autonomous agent throughput enables engineering teams to scale quarterly output from 3 to 300 initiatives, forcing organizations to re-evaluate their true core competencies rather than focusing solely on execution volume. (6:31)

Detailed notes

## Background Published on YouTube by the official Claude channel on October 23, 2026, this video features a discussion between Anthropic's Kai Gill and Factory Co-founder and CEO Matan Grinberg. According to the page description, the conversation explores Grinberg's journey founding Factory, how Claude's frontier models enable autonomous AI coding agents ('Droids') to execute long-horizon enterprise tasks with minimal human oversight, and what is required to build a fully autonomous 'dark factory' for software.

## TL;DR Matan Grinberg details Factory's mission to bring full autonomy to software engineering using multi-agent 'Droids' powered by frontier models like Claude Opus. After transitioning from theoretical physics research into enterprise software, Grinberg weathered early market timing hurdles to develop autonomous workflows that dramatically reduce human toil—such as reducing manual effort in enterprise COBOL modernizations by 90%. By assigning top-tier reasoning models to high-level architecture and planning, developers can build trust in agent-generated blueprints without micromanaging every step. Looking forward, Factory envisions a 'dark factory' paradigm where autonomous systems monitor, sandbox, and remediate production code around the clock, freeing engineers to focus on higher-leverage creative challenges.

## Atmosphere & style The conversation has an intellectually curious, candid, and collegial tone, punctuated by dry humor about Grinberg's unconventional pivot from 10-dimensional black holes into messy legacy codebases. The dialogue balances pragmatic engineering realities with an audacious vision, shifting naturally from self-deprecating reflections on early operational naiveite to touching anecdotes of enterprise engineers weeping with relief over automated migrations. It feels less like a polished corporate pitch and more like a thoughtful, philosophical fireside chat between peers navigating the frontier of autonomous systems.

## Bulletin - [05:29](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=329s) Factory leverages Claude Opus to conduct long-horizon architecture and planning for autonomous coding agents ('Droids'), enabling enterprise software developers to trust generated plans with minimal oversight. - [04:09](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=249s) Deploying Droids into a major bank for enterprise COBOL modernization reduced required human involvement by 90% after the initial onboarding pass. - [08:12](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=492s) Factory conceptualizes software autonomy as three concentric circles—frontier model, agent harness, and enterprise workflow—warning that builders must be ready to tear down orchestration layers monthly as models continuously absorb scaffolding functions. - [06:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=391s) Autonomous agent throughput enables engineering teams to scale quarterly output from 3 to 300 initiatives, forcing organizations to re-evaluate their true core competencies rather than focusing solely on execution volume.

## Key takeaways - [00:19](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=19s) Pivoting from 12 years of theoretical physics studying string theory and 10-dimensional black holes, Matan Grinberg founded Factory in 2023 after recognizing code generation's commercial potential in enterprise environments. - [01:48](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=108s) Factory's founding mission has remained steadfast: bringing complete autonomy to enterprise software engineering to free developers from low-leverage toil. - [02:22](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=142s) Operating ahead of underlying model capabilities between 2023 and 2025 gave Factory a critical runway to master operational and enterprise execution before market demand surged. - [02:58](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=178s) The 'dark factory' paradigm envisions autonomous systems running 24/7 without intervention, detecting production degradations and testing fixes in isolated sandboxes before alerting developers. - [02:58](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=178s) Real-world autonomous agents must adapt flexibly to the unique workflows and intelligence of enterprise developers rather than imposing rigid, top-down workflows. - [04:09](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=249s) In a major banking deployment, Droids learned domain-specific modernization rules alongside human engineers, reducing manual overhead on subsequent COBOL passes by 90%. - [04:58](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=298s) Multi-agent systems require hierarchical intelligence allocation: top-tier models like Claude Opus manage system-wide architecture and planning, while routine code execution is delegated downstream. - [05:29](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=329s) Opus-class reasoning shifts developer psychology from micromanaging step-by-step plans to defaulting to trust, preventing catastrophic compounding errors across multi-million-token runs. - [06:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=391s) Massive increases in feature deployment capacity compel engineering leaders to shift focus from raw feature delivery to defining and refining core organizational value. - [07:11](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=431s) Factory and Anthropic prioritize immediate agentic safety, alignment, and sandboxing as critical operational necessities rather than speculative future challenges. - [08:12](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=492s) Agent builders must stay structurally agile and frequently discard bespoke harness code, as advancing base models continually absorb capabilities once requiring manual scaffolding. - [09:18](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=558s) Visibility into frontier lab roadmaps allows agent developers to avoid wasting effort on transient reasoning limitations that base models will solve in upcoming iterations. - [09:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=571s) Aspiring founders should be guided purely by consuming intellectual obsession rather than seeking outside validation or permission to launch a startup.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=0s) Rethinking Human Work with Frontier Models

Matan Grinberg opens by reflecting on how technological leaps reshape human creativity and productivity. Throughout history, the most profoundly human pieces of art and technology have come from completely rethinking previous paradigms and uncovering entirely new angles of approach. With frontier models taking on repetitive, time-draining tasks, builders and creators are finally afforded the mental space and time required to step back, reconsider their methods, and tackle higher-order creative and systemic challenges.

[00:19](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=19s) From String Theory and 10-Dimensional Black Holes to Founding Factory

Kai Gill introduces Matan Grinberg, pointing out his remarkably non-traditional path into tech entrepreneurship. Until early 2023, Matan was an academic deep in the weeds of theoretical physics, researching the AdS/CFT correspondence in 10-dimensional black holes and string theory. Matan jokes that growing up in the Bay Area, startup culture was omnipresent, leading him to rebel by pursuing the furthest possible extreme: 12 years of theoretical physics. As Gill quips, “As one does,” Matan agrees, noting the realization that precipitated his shift between March and April 2023: theoretical physics and string theory were mathematically gorgeous, but we neither live in 10 dimensions nor anywhere near a black hole.

The turning point came when Matan took a seminar in program synthesis—an academic field now commonly referred to as code generation. The subject completely nerd-sniped him, leaving him head-over-heels obsessed with the potential of automating code. He quickly recognized that academia was not the environment to crack this problem. To build real code generation systems, one had to confront the messy reality of enterprise software: 30-year-old legacy codebases, systems half-written in COBOL and half in Python, and constantly evolving dependencies.

- **Background:** Matan spent roughly 12 years in theoretical physics studying high-energy theory, string theory, and AdS/CFT duality before pivoting to AI. - **The Transition:** The dividing line fell squarely between March and April 2023, when an academic seminar on program synthesis redirected his focus toward industry software engineering.

[01:48](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=108s) The Mission: Bringing Autonomy to Software Engineering

Factory was founded with a singular objective that has remained unchanged since its inception. Matan emphasizes that the very first pitch deck they created laid out the exact same core thesis they pursue today: bringing true autonomy to software engineering. Software engineers represent some of the most capable problem solvers on the planet, yet within large enterprises, an enormous portion of their daily work is bogged down by low-leverage toil, boilerplate changes, and legacy maintenance rather than high-leverage architectural problem-solving.

When Gill asks about having the conviction to build an autonomous software factory back when underlying foundational models were far from capable of handling it, Matan explains that they bet early on continuous model scaling. They anticipated that models would steadily advance to absorb more complex tasks over time. As Matan reflects on that transition: “Before, we were selling the vision of agents, now it's like everyone knows.”

[02:22](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=142s) The Journey Through the Desert: Lessons from Being Too Early

The two-year stretch between 2023 and 2025 served as Factory’s grueling “journey through the desert.” Gill observes that being too early in a market is often practically indistinguishable from being flat-out wrong, prompting Matan to share what that period of early isolation taught him. Factory was Matan's first real job; he was entirely obsessed with problem-solving but admitted to knowing virtually nothing about running an actual enterprise—from hiring and product design to sales, marketing, and organizational management.

Matan looks back at the timing mismatch as a critical blessing in disguise. Had the broader market and model capabilities been fully ready in early 2023, the startup would likely have collapsed under his lack of operational experience. As Matan puts it: “If the timing was right and the world was ready, I was the wrong founder in my mind because I didn't know anything.” That desert period provided a two-year training ground where he could learn how to build, operate, and scale a company right as the foundation models matured to match their original product ambitions.

[02:58](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=178s) Droids and the "Dark Factory" for Software

The conversation turns to Factory’s flagship product—autonomous agents called "Droids"—and the guiding vision of the "dark factory," a concept borrowed from industrial manufacturing where automated plants operate completely unattended with the lights turned off 24/7. Matan explains that achieving this in software requires gradually rewiring developer psychology. When developers first delegate long-horizon tasks that run for days or weeks without direct intervention, they must learn to surrender constant micro-management. As the Droids repeatedly deliver meaningful, correct code, developers begin developing the necessary trust to leave them unattended.

In this dark factory paradigm, an autonomous software factory continuously monitors production health, identifies degradations, tests fixes in isolated sandboxes, and proactively alerts engineers with verified solutions rather than waiting for humans to spot bugs. Matan emphasizes that achieving this requires deep respect for enterprise developers rather than attempting to enforce a rigid, one-size-fits-all workflow. As Matan insists: “I think it's so important to meet our users where they are and respect their intelligence as developers.” The platform must be malleable enough to bend to the silhouette and unique practices of each customer's organization.

```mermaid flowchart TD accTitle: Autonomous Dark Factory Loop accDescr: Flowchart illustrating how Factory Droids continuously monitor, sandbox, verify, and resolve production software issues autonomously.

subgraph Production ["Production Environment"] Deploy["Deployed Service"] Monitor["Continuous Telemetry & Monitoring"] Deploy -->|"metrics drift"| Monitor end

subgraph AutonomousLoop ["Autonomous Droid Remediation"] Detect["Detect Degradation\ne.g., completion rate drop"] Sandbox["Spin Up Test Sandbox"] Diagnose["Isolate Root Cause & Generate Candidate Fix"] Validate["Run Automated Verification"]

Monitor --> Detect Detect --> Sandbox Sandbox --> Diagnose Diagnose --> Validate end

subgraph HumanResolution ["Review & Deployment"] Notify["Ping Developer with Candidate Solution"] Apply["Deploy Validated Patch"]

Validate --> Notify Notify --> Apply Apply --> Deploy end ```

[04:09](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=249s) Modernizing Legacy COBOL with 90% Less Human Effort

When Gill asks for a standout real-world example of Droids executing complex end-to-end work, Matan points to enterprise banking, where nearly every institution maintains a massive backlog of legacy COBOL migrations that human engineers dread. Being assigned to manual COBOL modernizations is widely viewed by engineers as a grueling year-long sentence of soul-crushing toil. Factory deployed Droids into one major bank to tackle these modernization workflows directly alongside customer teams.

During the initial pass, Factory's engineers sat side-by-side with the bank’s internal developers while the Droid executed the migration, extracting domain-specific patterns and learning organizational rules as it worked. Armed with that learned context, the subsequent modernization pass achieved an astonishing reduction in manual overhead. As Matan recounts: “Then after that, the next one, the amount of human involvement was 90% less.” The breakthrough prompted emotional reactions from staff who had faced endless manual migrations, resulting in genuine tears of joy among engineers relieved of those burdens.

[04:58](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=298s) Strategic Architecture: When Droids Require Frontier Models

Gill and Matan discuss the structural allocation of intelligence within multi-agent systems and where frontier-class reasoning is essential. Matan draws an analogy to human engineering organizations: having a company's CTO write every single routine line of code is an inefficient waste of executive talent. Instead, engineering hierarchies rely on seasoned architects with deep institutional context to map out high-level system blueprints and coordinate dependencies, while delegating localized implementation to others.

Factory structures Droid workflows similarly, reserving frontier models like Claude Opus for overarching system architecture, long-horizon planning, and complex multi-step orchestration. When managing intricate codebases with dozens of moving parts, utilizing top-tier reasoning engines prevents cascading architectural errors that compound downstream across millions of generated tokens.

[05:29](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=329s) What Opus Unlocks: High-Trust Planning and Reduced Human Oversight

Discussing the concrete operational shift enabled by Claude Opus, Matan highlights the drastic reduction in human supervision required during complex tasks. In long-horizon agent execution, faulty initial planning is disastrous because a flawed strategy often only surfaces after consuming billions of tokens across downstream tasks. With Opus-grade reasoning, the planning phase is sound enough that human engineers no longer need to scrutinize every step before giving the green light.

Developers transition from tightly auditing every intermediate step to trusting agent blueprints by default. As Matan explains the qualitative shift in engineer trust: “I think there's the gut sense of you can trust the plans a little bit more to the point where you literally might not even look at them, and you're like, yeah, sounds great.” It has reached a state where only rare, bizarre corner cases prompt an engineer to pause and question the underlying plan, reliably clearing the strict thresholds demanded by enterprise environments.

[06:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=391s) From Shipping 3 to 300: Rethinking Core Competencies

Frontier AI removes the execution bottleneck that historically constrained engineering roadmaps. Matan notes that most companies enter a quarter hoping to ship 30 major initiatives, but resource limits restrict them to delivering three. With autonomous agents absorbing the sheer volume of implementation toil, an engineering organization suddenly possesses the capacity to ship 300 things in the same timeframe.

However, this explosive increase in capacity forces a fundamental identity question: does shipping 100 times more features actually move the needle for users and build stronger customer relationships? By taking over routine production burdens, models push human teams to re-evaluate their true leverage. As Matan observes: “I think it shifts the question for the humans to now be like, what really is our core competency? How can we deliver on that better?” Freeing teams from low-level execution creates space to invent completely new angles of attack and craft high-value software experiences.

[07:11](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=431s) The Anthropic Partnership and Shared Safety Priorities

Matan discusses Factory’s partnership with Anthropic, emphasizing that technical reliability and philosophical alignment are paramount. Both companies share an ambition to dramatically expand human potential while maintaining rigorous safety standards, clear messaging, and robust operational guardrails. Advanced AI capabilities carry immense power, which can naturally feel daunting to developers and enterprise leadership, making it vital to demonstrate concrete, grounded benefits while proving how Droids and Claude strengthen each other.

Matan stresses that agentic safety, alignment, and reliability cannot be treated as speculative future topics. As Matan emphasizes: “It's really important that the people that we partner with take it as seriously as we do, as a today problem.” The complexities of deployment, sandboxing, and autonomous code execution require frontier AI labs and tooling platforms to treat engineering constraints with immediate urgency.

[08:12](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=492s) The Three Concentric Circles Blocking the Dark Factory

Looking toward the near future, Gill asks what bottlenecks still prevent the full realization of lights-out dark factories in software engineering. Matan conceptualizes the current technical frontier as three rapidly expanding concentric circles: the foundational model at the center, the agent harness surrounding it, and the overarching company workflows on the perimeter.

```mermaid flowchart LR accTitle: The Three Concentric Circles of Software Autonomy accDescr: Diagram showing the inner foundation model, the middle harness orchestration layer, and the outer enterprise workflow layer.

subgraph Outer ["Outer Layer: Enterprise Ecosystem"] direction TB Workflows["Company Workflows & Cultural Silhouettes"] subgraph Middle ["Middle Layer: Agent Architecture"] direction TB Harness["Agent Harness, Tooling & Orchestration"] subgraph Inner ["Core Layer: Intelligence"] Model["Frontier Foundation Model\ne.g., Claude Opus"] end end end ```

The capabilities of standalone models have advanced so rapidly that a modern model can execute single-handedly what required a complex model-plus-harness configuration just two years prior. As frontier models absorb capabilities previously hardcoded into agent scaffolding, tooling builders must constantly re-evaluate their architectures rather than stubbornly defending outdated code. As Matan warns: “We need to always have in one hand a sledgehammer to whatever we thought was the best solution. Every month, we need to rethink.” Builders must continuously rethink whether their orchestration harnesses remain optimal or if rising model intelligence has rendered them obsolete.

[09:18](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=558s) Riding the Frontier Model Wave

Matan Grinberg explains the strategic value of maintaining a close relationship with Anthropic, emphasizing how visibility into the model roadmap shapes Factory's engineering roadmap. Because Factory can see the specific classes of challenges frontier lab researchers are tackling, the team can anticipate where base model capabilities will land in the near future. Rather than attempting to solve foundational reasoning hurdles that foundational models will soon make obsolete, Factory focuses on building the scaffolding and infrastructure that bridges current gaps. As Matan Grinberg puts it: “Let's make sure we're filling this hole in the meantime, so that we can ride that wave together.”

[09:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=571s) Advice for Would-Be Founders: Obsession and Intellectual Curiosity

Bringing the discussion full circle back to the pivotal moments that launch entrepreneurial journeys, Kai Gill asks what guidance Matan Grinberg would offer to aspiring founders watching the conversation who are standing at the threshold of starting something new. Grinberg responds with an immediate, humorous contrarian rule: “This is a bit funny, but anytime a person asks me if they should start a company, I always say no.” His rationale is straightforward: if an entrepreneur still needs external permission or affirmation to take the leap, they are not yet ready for the grueling realities of company building.

Instead, Grinberg argues that starting a company must be powered by deep, intrinsic fascination with a problem space. When a founder is genuinely absorbed by what they are building, the day-to-day grind transforms into something enjoyable. He encourages founders to pinpoint an obsession so consuming that it dominates their thoughts: “Find the thing that you literally cannot stop thinking about and just keeps you up at night.” Furthermore, Grinberg highlights that the technological landscape has drastically shifted the boundaries of feasibility. Whereas aiming for an extreme ambition like becoming an astronaut fifty years ago might have been practically unreachable, today's environment makes even the most radical concepts unexpectedly achievable. He advises would-be founders that as long as intellectual curiosity pours out naturally, everything else needed to build the company will follow.

- **Founder self-selection:** Anyone seeking outside validation or asking for permission to start a company is likely not ready to endure the journey. - **The modern technological era:** Extremely ambitious, unconventional ideas that were previously deemed impossible are now vastly more achievable. - **Guiding principle:** “Find the thing that you get obsessed about, the intellectual curiosity just pours out of you, and I think the rest will come from that,” Matan Grinberg concludes before wrapping up the conversation with Kai Gill.

## Notable quotes > [08:12](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=492s) We need to always have in one hand a sledgehammer to whatever we thought was the best solution. Every month, we need to rethink. — Matan Grinberg

> [09:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=571s) Anytime a person asks me if they should start a company, I always say no. — Matan Grinberg

> [02:58](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=178s) I think it's so important to meet our users where they are and respect their intelligence as developers. — Matan Grinberg

> [09:31](https://www.youtube.com/watch?v=N6FkmdL0Lg0&t=571s) Find the thing that you literally cannot stop thinking about and just keeps you up at night. — Matan Grinberg

## Mentioned Factory, Claude, Claude Opus, Anthropic, Matan Grinberg, Kai Gill, Droids, COBOL, Python

## Video description Factory is bringing autonomy to software engineering with Droids, AI coding agents that handle enterprise work. In this conversation, Co-founder and CEO Matan Grinberg sits down with Anthropic's Kai Gill to talk about his journey founding Factory, how Claude's frontier models handle long-horizon tasks across large codebases so teams can trust agent plans with less oversight, and what it will take to build a fully autonomous "dark factory" for software.

Read the full story: http://claude.com/customers/factory

Chapters: 0:00 Intro 0:19 From string theory to founding Factory 1:48 The mission: autonomy for software engineering 2:22 Lessons from being too early 2:58 Droids and the "dark factory" 4:09 A COBOL modernization with 90% less human effort 4:58 When Droids need frontier models 5:29 What Opus unlocks: less human oversight 6:31 Shipping 300 things and rethinking what matters 7:11 The Anthropic partnership 8:12 What's blocking the dark factory 9:31 Advice for would-be founders

[Source](https://www.youtube.com/watch?v=N6FkmdL0Lg0)

中文简报

Factory部署Claude Opus，削减企业90%编码工作量前沿模型可靠的远期规划能力消除了开发者的微观管理，随着基础能力演进，迫使软件团队不断淘汰自定义智能体脚手架。 Matan Grinberg详细介绍了Factory利用由Claude Opus等前沿模型驱动的多智能体“Droid”，实现软件工程全面自动化的使命。在从理论物理研究转向企业软件领域后，Grinberg克服了早期的市场时机难题，开发出可大幅减少人工辛劳的自动化工作流——例如在企业COBOL现代化改造中将手动工作量减少90%。通过将顶级推理模型分配给高层架构和规划，开发者无需微观管理每一步，即可对智能体生成的蓝图建立信任。展望未来，Factory构想了一种“黑灯工厂”范式，自动化系统在此模式下全天候监控、沙箱测试并修复生产代码，让工程师得以专注于更高杠杆的创造性挑战。 Factory利用Claude Opus为自主编码智能体（“Droid”）执行远期架构和规划，使企业软件开发者能够以极少的监督信任生成的方案。 (5:29) 在一家大型银行部署Droid进行企业COBOL现代化改造后，初次入职磨合期过后所需的人工参与减少了90%。 (4:09) Factory将软件自治构想为三个同心圆——前沿模型、智能体治理载体（harness）和企业工作流，并警告开发者必须准备好每月拆除编排层，因为模型正不断吸收脚手架功能。 (8:12) 自主智能体的吞吐量使工程团队能够将季度产出从3个项目扩展到300个项目，促使企业重新评估其真正的核心竞争力，而不仅仅是关注执行体量。 (6:31)

Reading summary

Factory founder Matan Grinberg argues that advances in frontier models allow software engineering to transition toward autonomous, unattended operations, freeing engineers from repetitive implementation toil to focus on core systemic problems. From Theoretical Physics to Autonomous Code Generation (0:19) Matan Grinberg transitioned from a 12-year academic background in theoretical physics—studying string theory and 10-dimensional black holes—into software entrepreneurship between March and April 2023. His focus shifted after a seminar on program synthesis convinced him that automating code required confronting messy enterprise realities, including decades-old legacy systems spanning COBOL and Python, rather than working within academic abstractions. Founding Factory on the thesis of full autonomy for software engineering, Grinberg initially faced a capability gap in base models between 2023 and 2025. He views this early period as essential, as it allowed him to develop core operational, hiring, and sales skills before foundational models matured enough to realize Factory's product goals. The Dark Factory Model and Legacy Modernization (2:58) Factory structures its autonomous agents, termed 'Droids,' around the manufacturing concept of a 'dark factory'—environments designed to operate continuously without human presence. In software, this model involves autonomous systems that monitor production telemetry, detect degradations, reproduce issues in sandboxes, and present verified candidate patches to developers rather than waiting for manual triage. Grinberg notes that achieving this requires shifting developer psychology from micro-management to trusting unattended execution over multi-day horizons. In practical enterprise deployments, such as legacy modernization inside major financial institutions, this approach has demonstrated measurable reductions in human effort. Grinberg reports that after Droids paired with bank engineers on an initial COBOL migration to capture organizational context, the subsequent modernization pass required 90% less human involvement, eliminating major backlogs of manual migration work. Tiered Intelligence and Claude Opus Planning (4:58) Factory approaches multi-agent system design hierarchically, drawing a comparison to human engineering organizations where senior architects design overarching blueprints rather than writing every routine line of code. Droid workflows delegate routine localized implementation while reserving frontier reasoning models like Claude Opus for long-horizon planning, system architecture, and orchestrating multi-step dependencies across codebases. Grinberg states that incorporating Opus-level reasoning drastically reduces supervision requirements. Because flawed plans in long-horizon agent workflows can silently consume vast token budgets before failing, higher-grade initial planning allows engineers to approve architectural blueprints by default, pausing only when encountering edge cases. Adapting Scaffolding to Expanding Model Capabilities (8:12) Grinberg describes software autonomy as three concentric layers: the core foundation model, the intermediary agent harness and tooling, and the outer layer of enterprise workflows. Because standalone models advance rapidly—frequently absorbing tasks that previously required complex external scaffolding—tooling developers must continuously reassess and dismantle obsolete harness code rather than defending outdated architectures. Through its partnership with Anthropic, Factory aligns its roadmap with upcoming frontier capabilities. Rather than building complex temporary workarounds for foundational reasoning limitations that labs are already solving, Factory focuses on infrastructure, safety guardrails, and sandboxed deployment environments that bridge immediate operational gaps. While autonomous agents allow organizations to scale output significantly, Grinberg observes that removing execution bottlenecks forces enterprises to re-evaluate their true core competencies and customer value. Realizing fully autonomous software engineering remains constrained by how quickly developer trust forms, safety guardrails evolve, and agent harnesses adapt to advancing base models.

精炼阅读摘要

Factory 创始人 Matan Grinberg 认为，随着前沿模型接管重复性的低杠杆开发工作，软件工程正迈向无需人工时刻干预的“黑灯工厂”自主模式，促使开发团队将重心转向高阶架构设计与核心竞争力定义。从理论物理转向软件工程的自主化使命 (0:19) Matan Grinberg 在从事了约 12 年的高能物理、弦理论及十维黑洞研究后，于 2023 年 3 月至 4 月间参与了一场关于程序合成的研讨会，随后决定转向产业界创办 Factory。他认为，虽然学术理论具备数学美感，但真实的工程自动化必须直面包含 30 年陈旧系统、混杂 COBOL 与 Python 等复杂依赖的企业现状。Factory 自创立之初即确立了实现软件工程真正自主化的目标，将软件工程师从低效的模板化编码与遗留维护中解放出来。软件“黑灯工厂”构想与开发信任建立 (2:58) Factory 借鉴了工业制造中全天候无人值守的“黑灯工厂”理念，推出了名为 Droids 的自主智能体。在设想的运行闭环中，自主系统能持续监控生产环境健康状态、在独立沙盒中复现并修复问题，并向工程师主动提供经过验证的补丁。Grinberg 指出，实现这一愿景的核心在于重构开发者的心理认知，使他们逐步放弃微观管理。在某大型银行的历史 COBOL 系统迁移实践中，Droid 在经过初期与人工的协作学习后，后续环节的人工介入程度降低了 90%。分层智能架构与前沿规划模型的关键作用 (4:58) 在智能体架构设计上，Factory 采取了类似企业工程梯队的智能分配策略，将 Claude Opus 等前沿大模型定位为负责宏观系统蓝图、跨周期规划与复杂多步骤编排的架构师，以避免在庞大代码库中因顶层错误引发下游巨量 Token 的连带崩塌。具备高阶推理能力的模型显著提升了规划阶段的可靠性，使工程师无需逐一审核每个中间步骤，基本确立了“默认信任”的工作方式，唯有遇到罕见边缘情况时才需介入干预。交付能力剧增与企业核心定位的再思考 (6:31) 前沿人工智能打破了过去受人力瓶颈约束的交付规划，使企业团队在一个季度内完成原本难以企及的任务量，例如将仅能推进 3 个项目的产能提升至能够交付 300 项任务。然而 Grinberg 强调，产能的大幅释放迫使人类团队重新审视自身的真实杠杆与核心专长，思考单纯的数量倍增是否真正有助于加深客户关系与产品价值，从而倒逼团队探索更高阶的软件体验与创新方法。三层同心圆演进与智能体基础设施的快速迭代 (8:12) 针对迈向黑灯工厂的阻碍，Grinberg 提出了由基础模型、智能体 Harness 套件以及企业工作流构成的三层同心圆框架。随着底层基础模型能力每两年出现质的飞跃，过去需要复杂外部 Harness 拼凑实现的功能正逐步被独立模型直接原生覆盖。因此，工具构建者必须保持灵活的技术态度，每月重新评估既有架构，避免固步自封；通过与 Anthropic 等实验室保持路线图协同，重点构建填补当前能力空缺的过渡设施，动态顺应前沿模型演进浪潮。自主软件工程的发展依赖于底层前沿模型演进与工程架构的敏捷自省，但要完全普及仍需克服企业个性化工作流的适配挑战以及重塑开发者对长时间无人值守系统的长期信任。

Source quotes

[8:12] Matan Grinberg: We need to always have in one hand a sledgehammer to whatever we thought was the best solution. Every month, we need to rethink. [9:31] Matan Grinberg: Anytime a person asks me if they should start a company, I always say no. [2:58] Matan Grinberg: I think it's so important to meet our users where they are and respect their intelligence as developers. [9:31] Matan Grinberg: Find the thing that you literally cannot stop thinking about and just keeps you up at night.

Original transcript

[0:00] I think that's some of the most human [0:03] pieces of art or pieces of technology, rethink what we've been doing before [0:07] and figure out this entirely new angle. [0:09] And we have time to do that by using models like Opus and Fable, [0:12] where they take on a lot of the work that used to occupy so much of our time. [0:19] Matan, super great to be here with you today. [0:21] We were talking about this a little bit before we sat down, but you [0:24] have such an unusual background. [0:26] In 2023, there's this pivotal moment where on one side of it, you were a PhD student [0:31] studying the AdS/CFT correspondence in 10-dimensional black [0:34] holes and string theory. [0:36] On the other side of it, you become a founder, of Factory. [0:39] What happened? [0:41] I guess growing up in the Bay Area, the culture is startups. [0:47] And I was a rebellious child. [0:49] I was like, I want nothing to do with startups. [0:50] I'm going to do the most opposite thing. I got into theoretical physics [0:54] and did that for about 12 years. As one does. [0:57] Yeah, right. [0:59] That dividing line between March and April 2023, I think that was the culmination [1:03] of realizing that while theoretical physics and string theory is extremely [1:07] beautiful, as you've noticed, I was working on 10-dimensional black [1:11] holes, and we don't live in 10 dimensions, nor do we live near a black hole. [1:15] But then I took a seminar in program synthesis, which we now call code [1:18] generation, and that just completely nerd-sniped me. [1:21] Head over heels, was obsessed with that. [1:23] Then it just became very clear in 2023 that to solve this problem, [1:27] academia was not the place. [1:28] You had to go into the industry. [1:30] You had to work with some of the world's largest enterprises, [1:33] some of the messiest codebases where some of the code is 30 years old, [1:36] some of it is brand new, some of it is in COBOL, half of it is in Python. [1:41] Just this very weird dynamic situation is where the code generation problem [1:47] becomes the most interesting. [1:48] The mission that we have in the very first deck that we ever made [1:52] is the same as our mission today, which is to bring autonomy to software engineering. [1:56] Software engineers are some of the best problem solvers in the world, [2:00] but so much of their life, especially in a large enterprise, [2:03] is not as high leverage as it could be. [2:06] You had this conviction from day one that this is going to be an autonomous [2:10] software factory, even though at the time, the models weren't really there yet. [2:14] Models are going to get better. [2:15] They're going to be able to do more and more of these tasks. [2:18] Before, we were selling the vision of agents, now it's like everyone knows. [2:22] That 2-year period between 2023-2025 was your journey through the desert. [2:26] I think you said elsewhere that sometimes being too early is [2:29] the same as being wrong. What did that desert teach you? [2:31] I was obsessed with problems and solving problems. [2:34] That was all I cared about. [2:35] This is really the first job I've ever had. [2:37] I knew nothing about anything relating to building a company. [2:40] Whether it's hiring, engineering, building product, sales, marketing, none of it. [2:45] If the timing was right and the world was ready, I was the wrong founder in my mind [2:49] because I didn't know anything. We would have missed the opportunity. [2:51] It's kind of a blessing in disguise because I think it allowed me the time [2:55] to grow and learn the things that really matter to now scale. [2:58] You've mentioned this concept of these Droids working in a dark factory, [3:02] lights off 24 hours a day, 7 days a week. [3:05] How are you building the product now for that future? [3:10] As you get used to delegating tasks to agents that can go for days or weeks [3:14] on end, you start getting used to the idea that there's something that you're not [3:18] going to go and touch for a couple of days at a time. [3:20] It slowly starts building that behavior change. [3:22] Also, when they see that the output is actually meaningful, they start [3:26] to build some of that trust. [3:27] Trust is going to become all the more important. [3:29] You're now having systems that are doing things without you looking, [3:31] without you necessarily explicitly approving every time. [3:34] Having a software factory that is always monitoring, seeing, oh, hey, [3:38] we just deployed some change, and now the completion rate [3:42] through this step has gone down. [3:43] Let's go and spin up a sandbox where we go and test and see why did that happen? [3:48] We found a candidate solution. [3:50] Let me go ping you and show you what might fix it, and then deploy it and monitor [3:54] again. I think it's so important to meet our users where they are and respect [3:59] their intelligence as developers. [4:00] They know their business more than we do. [4:03] And the best we can do is allow them to configure it and morph it [4:07] to the silhouette of their organization. [4:09] Do you have any hero examples of a Droid doing something really impressive [4:14] or crazy end-to-end for a customer? [4:16] Every single bank has a backlog list of these crazy COBOL modernizations [4:22] that they need to do. [4:23] It's all like, no engineer wants to be assigned to this because that is like, [4:26] "Oh, God, for the next year, my life is going to suck dealing with this." [4:29] With one such bank, we did one of these modernizations. [4:32] As it was working, it was extracting lessons. [4:35] This is the way this bank does things. [4:37] This is how we can learn from it in the future. [4:39] That was one of our engineers sitting next to some of their engineers [4:42] to go and do that first pass. [4:44] Then after that, the next one, the amount of human involvement was 90% less. [4:49] That to them, was a magic moment. [4:51] There are tears of joy in these engineers' eyes because working [4:55] on those migrations is the worst. [4:58] Where do you feel that the Droids need to be very, very smart, [5:03] and they need the frontier? [5:05] If you have the CTO of a company implement every line of code, that is not [5:09] an efficient use of their time. [5:11] It turns out that typically we will have structure where there is someone who has [5:14] maybe more context about a problem or more history of problem-solving, [5:18] where they're going to come up with a general systems' solution. [5:21] It's analogous here where, I myself, if I'm doing planning or something pretty [5:26] complicated where I need to orchestrate a lot of moving parts, I'll use Opus. [5:30] What has Opus enabled you to build with the Droids that you [5:33] weren't able to before? [5:34] I think the biggest unlock, and this is the most important part [5:37] for us, is reducing human oversight. [5:40] That planning is a lot more difficult, or you'll only find out if you went awry, [5:45] billions of tokens down the line. [5:47] I think there's the gut sense of you can trust the plans a little bit more [5:53] to the point where you literally might not even look at them, and you're like, [5:55] yeah, sounds great. [5:56] Let's go and do it. That's been pretty cool to see. [5:58] Before, it was like, you see it, but you're really closely [6:00] monitoring to now it's like… [6:02] It's only a very crazy corner case where you'll go back and be like, wait, [6:05] what did we do in the plan? It rarely, rarely happens. [6:08] There are certain things that are the brutal reality of working at a company [6:12] where there are just things that you have to do that need to meet [6:15] a certain reliability threshold. [6:17] And what's cool is Opus and Fable, it takes a lot of those things [6:21] that previously you had to spend a ridiculous amount of time doing, [6:24] and it's like, "I got you covered there." Every company wants to ship 30 things, [6:28] but they can only ship three and a quarter because of the resourcing that they have. [6:31] Well, now you can ship 300 things. [6:33] But will those 300 things actually make a difference? [6:36] Are they going to actually make happier customers or better relationships [6:39] with your customers or your end users? [6:41] I think it shifts the question for the humans to now be like, [6:45] what really is our core competency? [6:47] How can we deliver on that better? That is the thing that's really hard because [6:51] sometimes we put ourselves in such a box of like, this is all that we can do. [6:54] To me, it's very exciting because I think that's some of the most human [6:59] pieces of art or pieces of technology, rethink what we've been doing before, [7:03] and figure out this entirely new angle. [7:05] We have time to do that by using models like Opus and Fable, [7:08] where they take on a lot of the work that used to occupy so much of our time. [7:11] When you think of Anthropic and you think of that partnership, [7:14] what's top of mind for you? [7:15] I think the biggest thing is reliability in terms of how we work. [7:18] Understanding that there are a lot of things that we have in common in terms [7:22] of what we're excited about, and making sure that it's set up such [7:25] that like, hey, here are the important things we should spend our time on. [7:27] Here are the things that you guys should spend your time on. We very clearly want [7:32] to see a world where human potential is unlocked, and this ambition is raised, [7:38] and there are certain constraints that are required. [7:40] There're some safety constraints that are required. [7:42] There's some messaging that really matters. [7:44] Sometimes these tools can be scary. [7:46] They're very powerful, and that is inherently scary. [7:48] But it's also a beautiful thing for the world that we should highlight [7:52] the benefits that are coming from it and how Droid can be better with Claude [7:56] and how Claude can be better with Droid. [7:57] It's really important that the people that we partner with take it as [8:00] seriously as we do, as a today problem. [8:03] Some people are like, "Oh, eventually it might be crazy." It's like, [8:06] no, today this is extremely important across every piece of this [8:10] AI engineering puzzle. [8:12] Looking ahead a year in advance when you have these dark factories. [8:16] We're not fully there yet. [8:18] What are the blockers? [8:19] What are the bottlenecks right now? [8:21] The way I think about it, there are three concentric circles. [8:24] At the core, there's the model, then there's the harness, [8:27] and then there's the company and the workflows of the company as a whole. [8:31] These are all growing very quickly. [8:34] The model capabilities now versus 2 years ago, it's insane. [8:37] Completely, completely different. [8:38] A model today could do what a model plus a harness could do 2 years ago. [8:41] It's a cool call to arms for everyone to raise the scale of their ambition. [8:46] Because as models get better, if you are lazy, and you don't change [8:49] your harness, it's going to be superseded. [8:52] As these models get smarter and smarter, they start to absorb more of the harness. [8:56] How do you build for that? [8:58] Then also, how do you think about the leverage that that type of really [9:02] intelligent model can give you? We cannot settle. [9:05] Things are moving way too quickly. [9:06] We need to always have in one hand a sledgehammer to whatever we [9:10] thought was the best solution. [9:12] Every month, we need to rethink. [9:14] Is this still optimal? [9:15] As these models are getting better, and also having a deep partnership, [9:18] understanding and having this relationship where we can see, hey, [9:22] these are the types of problems they're thinking about. We can trust, okay, [9:25] this is going to be going up there. [9:27] Let's make sure we're filling this hole in the meantime, so that we [9:29] can ride that wave together. [9:31] I want to go back to the very beginning. [9:33] We were talking about your pivotal moment that led you onto this path. [9:37] There are a number of would-be founders watching this, and their pivotal [9:41] moment is right ahead of them. [9:43] What advice would you have to those founders? [9:45] This is a bit funny, but anytime a person asks me if they [9:47] should start a company, I always say no. [9:49] Because if they need my affirmation, you know. [9:52] I just like really getting obsessed with problems. [9:56] Then everything is just so much fun. [9:58] Find the thing that you literally cannot stop thinking about and just [10:01] keeps you up at night. [10:02] The beautiful thing about the era in which we're in now, [10:05] normally you could say, oh, the thing that keeps me up at night is [10:08] I want to be an astronaut, or whatever. [10:10] Fifty years ago, it's pretty unrealistic. [10:12] It is extremely realistic today. [10:14] Even if you have the most out there crazy ideas, and that's the only thing [10:19] that keeps you up at night, it's actually way more achievable than it's ever been. [10:22] Find the thing that you get obsessed about, the intellectual curiosity just [10:26] pours out of you, and I think the rest will come from that. [10:31] Thank you so much. This has been an amazing conversation. [10:34] I really appreciate the time. Thank you for having me. [10:36] This was fun.