Claude Code Manager: Under multi-agent parallelism, merge 20-30 PRs daily

Headline

The head of Claude Code at Anthropic is being inundated with AI-generated outputs.

Summary

Boris Cherny (the creator and head of Claude Code) mentioned on social media and podcasts:

• He receives a pile of AI-related outputs and notifications every day, mainly from the Claude agents he runs in parallel—code products, status alerts, and so on.
• Through multiple instances running in parallel, he manages to handle 20-30 PR submissions and merges each day.
• In interviews on Pragmatic Engineer and Lenny’s Podcast, he stated that this code is “100% written by AI,” and he hasn’t touched code himself since November.

The key point here is: This is a real engineering workflow driven by multi-agent collaboration, not a demonstration or a score run.

Analysis

• Work content has changed:

  • Planning, implementation, and review are executed in parallel by different Claude agents;
  • His responsibilities have shifted from “writing code” to “organizing and scheduling agents, accepting results”;
  • If what he says is true, this represents a significant change in the definition of daily work for engineers.

• Technical methods are surprisingly simple:

  • In large-scale codebase transformations (like migrations), Claude Code tends to use straightforward methods (like glob, grep), rather than complex and flashy techniques;
  • In practice, “simple tools + parallel organization” are sufficient to tackle complex tasks.

• Contradiction between organization and quality:

  • High-concurrency PRs pose risks to code quality, and systemic errors may be amplified;
  • He did not disclose quality thresholds and risk control details, but the boundaries of the issues are clear: if errors exhibit patterned characteristics, the degree of parallelism will magnify the impact.

Impact Assessment

• Importance: Medium-High (a high-frequency workflow sample from frontline practitioners).
• Paradigm shift:
  • Transitioning from “developers writing code by hand” to “multi-agent parallelism + human scheduling” in the pipeline;
  • Human engineers take on the role of overall control for products and processes, while AI is responsible for specific writing, modification, querying, and evaluation tasks.
• Scope of applicability:
  • Suitable for scenarios with high structural integrity that can be broken down into batch rewriting/migration/review tasks;
  • For complex tasks requiring strong interaction and cross-domain design decisions, human oversight for strategy and constraints is still necessary.

Conclusion (Quick Points)

• Source of production: Parallel agents + simple toolchain;
• Main risk: Systematic errors being amplified in parallel;
• Practical value: Real cases of frontline teams “eating their own dog food” are more persuasive than paper benchmarks.

Conclusion: For traders and funding parties, this narrative currently has limited relevance; for Builders who are building or introducing AI coding pipelines, this serves as an “early but effective” reference paradigm, with the main beneficiaries being those pursuing engineering efficiency and process orchestration.