Boris Cherny Uses Claude Code as an Engineering Lead: The 388-PR Experiment

Forum topic · 小凯 · 2026-08-14

Summary

In a Chinese tech forum discussion, users shared a talk and experiment attributed to Boris Cherny, creator of Claude Code, in which he treats the AI tool as an engineering team lead rather than a coding assistant. According to the shared material, Claude Code managed a workflow that produced 388 pull requests, with Cherny positioning the model as a manager that plans, delegates, and reviews work across multiple subagents instead of writing every change itself. The forum thread reflects strong interest in this “AI as tech lead” workflow, discussing how orchestrating parallel agents, giving the model architectural authority, and structuring tasks at the pull-request level can dramatically scale output. The post highlights a shift in how developers use AI coding tools: from autocomplete-style assistance to agentic systems that own entire delivery pipelines. It also raises questions about code review quality, supervision, and whether such throughput is sustainable in real production teams.

Overview

A forum thread on zhichai.net discusses an experiment by Boris Cherny, the creator of Claude Code, in which he uses the tool not as a pair-programming assistant but as an engineering lead responsible for managing work end to end.

Key Points

Cherny's setup treats Claude Code as a tech lead: it plans tasks, delegates implementation to subagents, and reviews results.
The experiment reportedly produced 388 pull requests, demonstrating throughput far beyond a typical human-led workflow.
The emphasis is on orchestration: rather than the model writing every line, it coordinates multiple parallel agents working at the pull-request level.
Forum participants see this as a signal of where AI coding tools are heading — from autocomplete and chat assistance toward fully agentic delivery pipelines.

Discussion Themes

Scalability: Breaking work into many small PRs lets parallel agents ship large volumes of changes quickly.
Supervision: With an AI acting as lead, humans shift from writing code to reviewing direction, architecture, and output quality.
Open questions: Commenters raise concerns about review depth, long-term maintainability, and whether such PR volume reflects genuine engineering value or inflated activity metrics.

Takeaway

The thread captures a notable mindset shift in AI-assisted development: instead of asking "how can AI help me code faster?", practitioners like Cherny are asking "how can AI run the engineering process itself?"