Boris Cherny, the engineer behind Anthropic's Claude Code, has a pointed message for developers who watch their AI agent's every move: back off. In a conversation with Business Insider, Cherny argued that the habit of micromanaging AI tools is one of the primary reasons people walk away disappointed. The fix, he says, is not a better model or a different prompt. It is learning to let go.
The Autonomy Argument
Cherny's position is straightforward. When developers interrupt an AI agent repeatedly to course-correct, redirect, or second-guess its approach, they break the workflow the tool was designed to sustain. Claude Code, built to handle long, multi-step coding tasks with minimal hand-holding, performs best when given a clear objective and then left to pursue it. As Cherny has noted before, he himself has not written a line of code in eight months, delegating that work almost entirely to the agent. That is not a stunt. It is a demonstration of the philosophy baked into the product.
Key Facts
- Boris Cherny is the creator of Claude Code at Anthropic
- He argues micromanaging AI agents actively reduces output quality
- Claude Code is designed for long-horizon, autonomous coding tasks
- Cherny has publicly said he delegates nearly all coding work to the tool
- The advice applies broadly to agentic AI workflows, not just coding
The broader point connects to a growing conversation in developer circles about how to work with agentic tools effectively. Many users approach AI agents the way they might approach a junior employee on probation, checking in constantly and overriding decisions before they play out. Cherny's view is that this posture transfers the cognitive load back to the human, which defeats the purpose. He has also pushed back on how developers measure agent performance, suggesting that token consumption is a poor proxy for actual productivity.
The people who get the most out of these tools are the ones who treat them like a capable colleague, not a tool they have to supervise constantly.Boris Cherny, creator of Claude Code, via Business Insider
What This Means for Everyday Users
The advice has practical implications for anyone using Claude Code in a professional setting. Rather than feeding the agent one small task at a time and waiting to approve each output, Cherny recommends scoping a larger goal and letting the agent navigate toward it. Errors will happen, he acknowledges, but interrupting the process to prevent every possible mistake often produces worse outcomes than letting the agent work through problems on its own terms.
This philosophy is increasingly central to how Anthropic positions its developer tools. The company has been building out infrastructure for agentic workflows, including managed environments where Claude can operate with persistent context over extended sessions. Developers exploring that ecosystem would benefit from understanding the setup and approach Cherny advocates, and a detailed look at Claude Code's recommended workflow offers a useful starting point for teams just getting started.
The cultural shift Cherny is describing is not trivial. Trusting an AI system to make a sequence of decisions without constant oversight requires a change in mindset that many developers are still working through. But the argument from results is hard to dismiss. The evidence from power users suggests that autonomy, given thoughtfully with a well-defined goal, tends to outperform the cautious, step-by-step approach that feels safer but often produces less.
As agentic AI tools mature and take on more complex tasks, the question of how much human oversight is actually helpful will only become more pressing. Cherny's answer, for now, is clear: less is more.
