Boris Cherny Ran Claude Code as an Engineering Lead: An Experiment with 388 Pull Requests

Forum topic · 小凯 · 2026-08-14

Summary

A post on zhichai.net discusses Boris Cherny's experiment treating Claude Code as an engineering lead rather than a coding assistant, in a workflow that produced 388 pull requests. The core idea is a management-style approach to AI coding agents: instead of reviewing and directing every line of code, the human sets direction, reviews outcomes, and delegates implementation tasks to the agent, similar to how a lead engineer manages contributors. This framing has drawn attention in the developer community because it shifts the mental model from pair programming to AI-driven delegation at scale. The discussion highlights practical implications for engineering teams considering agentic coding tools: workload division, review bottlenecks, trust in automated changes, and how metrics like merged PR counts should be interpreted. The post serves as a case study reference for developers evaluating Claude Code and similar AI coding workflows in real projects. Readers interested in agentic development, AI-assisted code review, and human-AI team structures will find the experiment details relevant to their own adoption decisions.

Overview

This forum post discusses Boris Cherny's experiment in using Claude Code not merely as a coding assistant, but as an *engineering lead* — delegating implementation work to the agent the way a tech lead delegates to engineers — with the experiment reportedly spanning 388 pull requests.

Key Points

Management-style delegation: Instead of pair-programming with the AI at the line level, Cherny's approach treats Claude Code as someone to whom tasks are assigned, with the human acting in a review/oversight capacity.
Scale of the experiment: 388 pull requests generated under this workflow, suggesting sustained, high-volume use of an agentic coding tool in a real project.
Implications for teams: The discussion raises questions about review bottlenecks, trust in agent-produced changes, and how to interpret PR counts as a productivity metric when agents automate implementation.

Takeaway

The experiment is a notable data point in the shift from AI-as-typist to AI-as-contributor workflows, and is relevant for teams evaluating Claude Code and similar agentic coding tools.

*Note: The original source content for this post was limited to its title; details above reflect only what is stated there.*