Vincent Koc, a core maintainer at OpenClaw, presents an insightful talk on how his team manages to ship code at an extraordinary velocity, often faster than one can read the diffs. He draws parallels between the current rapid development in AI and the Industrial Revolution, emphasizing that traditional ways of working no longer suffice. Vincent shares his personal journey with technology, highlighting his early experiences with VR and how working on cutting-edge projects often involves navigating uncharted and sometimes chaotic territory. He stresses that the incredible speed of development at OpenClaw is not mere luck but the result of deliberate engineering practices and process innovations.
The team at OpenClaw operates like a modern factory, with engineers acting as factory managers overseeing multiple “swim lanes” or parallel streams of work. These swim lanes allow them to manage different aspects of the codebase simultaneously, such as CI, features, bugs, and priority issues. Vincent explains that the bottleneck has shifted from manual craftsmanship to managing taste and quality at scale, especially as the volume of commits reaches thousands per day. He also discusses the challenges of handling such a high velocity of changes, including the need for a plugin architecture to modularize the codebase and prevent bloat, which they achieved through a massive refactor involving thousands of commits.
A key part of their success lies in the use of autonomous coding agents that assist with various tasks, running dozens of sessions concurrently. Vincent likens managing these agents to managing a team of people, where intuition and soft skills are crucial to detect when an agent is “bullshitting” or not performing as expected. He shares how he monitors the agents’ behavior and explanations to decide when to reset or reassign tasks, emphasizing that this human-in-the-loop approach is essential for maintaining quality and progress. The team also relies heavily on automated testing, which, despite being imperfect, provides a safety net during large-scale refactors.
Vincent outlines the engineering infrastructure supporting their workflow, including the use of “skills”—modular capabilities that agents can use to perform specific functions like writing documentation or running tests. These skills are managed and improved continuously, deployed across different environments to maintain efficiency. He also touches on the complexity of managing thousands of pull requests and issues, using semantic graphing and vector embeddings to cluster and prioritize work. This process helps the team focus on the most critical problems without getting overwhelmed by noise, illustrating the importance of structured workflows in high-velocity development environments.
In conclusion, Vincent highlights that the future of software development at scale is not just about maximizing token usage or raw compute power but about optimizing processes and human-agent collaboration. He predicts that 2025 was about pushing the limits of token consumption, while 2026 will focus on token efficiency and smarter agent-in-the-loop systems. Ultimately, the talk underscores that successful large-scale AI-driven development requires a blend of engineering rigor, intuitive management, and innovative tooling to harness the full potential of autonomous coding agents.
