Dark Factory: OpenClaw Ships Faster Than You Can Read the Diff — Vincent Koc, OpenClaw

Vincent Koc, core maintainer of OpenClaw, presents what he calls the "dark factory" model of AI-assisted software development, describing how his team achieved extraordinary shipping velocity using parallel autonomous coding agents. Koc opens by framing the current moment as analogous to the Industrial Revolution: just as handlooms gave way to centralized mills, individual engineers writing code in editors are giving way to engineers acting as factory managers overseeing swarms of agents across repositories. The bottleneck, he argues, has shifted from the weaver's hands to taste and judgment.

The talk's central case study is what Koc calls "the great refactor" — an overnight effort to decompose the OpenClaw monolith into a plugin architecture. Carried out while Koc and co-maintainer Peter Steinberger were physically at Nvidia building NemoClav, the refactor produced 2,700 commits, touched close to a million lines of code, and modified 82% of the core codebase in a single night. The plugin architecture — intended to let providers like OpenAI, Mistral, and Anthropic own their own sections of provider code — was shipped by morning. The saving grace for test stability was the overfitted unit tests that AI code generation tends to produce: because they over-indexed on implementation details, they served as a surprisingly reliable green-light signal even after wholesale structural changes.

Koc operates with what he calls swim lanes: 15 to 20 parallel coding sessions organized by work type. CI fixes and test refactors occupy low-supervision lanes where agents are told to commit and push autonomously. Feature work and specific bug lanes — for example around Docker integration or messaging channels — involve active back-and-forth conversation. A fifth category of lanes handles triage using live signals from Discord releases, GitHub issue clustering, and semantic vector embeddings over PRs. At the peak of the great refactor, Koc personally ran 10 to 15 foreground sessions with sub-agents pushing the effective total to 60 to 70 concurrent agents between himself and Steinberger.

The most technically nuanced section of the talk concerns what Koc calls "reading the reasoning tokens." He argues the skill in 2026 is not prompt engineering but knowing when an agent is hallucinating confidence. Drawing on the Matrix metaphor of reading the code cascade directly, he describes an intuitive pattern-matching ability built through sustained high-volume agent use: when an agent is waffling or explaining itself incoherently, the correct response is to nuke the session entirely rather than continue burning tokens. This intuition, he says, resembles managing people — experienced engineering managers learn to detect when a report is bluffing, and the same soft skill transfers directly to agent management.

Koc's infrastructure is deliberately simple. He manages skills using a dotfiles-style dot-skills directory, co-created with developer experience engineers and open-sourced on GitHub. He uses vercel.skills.sh as a skill registry and has built a self-healing Codex setup that handles the roughly 70 to 80 active Git work trees he accumulates daily. He explicitly does not use plan mode or spec mode, preferring direct conversational interaction with agents.

For PR management at the 60,000-PR scale OpenClaw faces, he built semantic graph clustering with vector embeddings over GitHub issues, representing a single PR as a node with 73 to 106 edges across related issues. The team also runs evaluation loops using a synthetic Slack environment populated with both synthetic and real model responses to validate that provider integrations and messaging channels continue working after refactors.

Koc closes with a macro thesis: 2025 was about token maximization — burning as many tokens as possible to produce output. 2026 is about token efficiency and agent-in-the-loop judgment. The constraints are no longer model capability or raw compute; they are the engineer's brain space and capacity to supervise the factory floor.