Boris Cherny runs thousands of Claude Code subagents
Claude Code
AI
Anthropic
TLDR: ASPENâBoris Cherny says he has not written code by hand for eight months, overseeing hundreds to tens of thousands of Claude Code subagents. The shift turns software work into managing AI production cycles, changing how developers build and how teams stay mentally sharp.
Key Takeaways:
- Boris Cherny leads Anthropicâs Claude Code, where AI agents generate and edit code from text briefs across large codebases.
- Cherny now manages âmaybe a few hundredâ agents, sometimes âthousandsâ up to âtens of thousands,â with Claude Code spawning prompts via subagents.
- He links the productivity leap to Gutenbergâs printing press, but warns that heavy AI oversight can trigger âAI brain fry,â including brain fog and slower decisions.
The job description has quietly flipped from typing to supervision, and that changes who feels in control. If AI can write the code, the real bottleneck becomes your ability to direct it without burning out.
The job description has quietly flipped from typing to supervision, and that changes who feels in control. If AI can write the code, the real bottleneck becomes your ability to direct it without burning out.
Q&A
If developers stop prompting Claude directly, what will âgood engineeringâ look like in day to day reviews?
It shifts toward writing sharper briefs, defining guardrails, and evaluating agent change summaries for correctness, safety, and maintainability.
How will teams measure quality when agents run overnight and produce large volumes of edits?
Expect tighter automated testing, stronger code review standards, and more structured acceptance criteria so human time focuses on high impact decisions.
What happens when agent workflows scale to tens of thousands of concurrent tasks?
Bottlenecks move to infrastructure, permissions, and resource limits, plus monitoring that can catch failures before they snowball across a codebase.
Why does a productivity boost also raise the risk of âAI brain fryâ for the humans involved?
Supervision requires constant context switching and judgment; when oversight exceeds cognitive limits, errors and slower decisions can rise even if the AI speeds up output.
Does the Gutenberg analogy imply that software will become more accessible, or just more chaotic?
It can democratize building, but only if organizations pair agent power with disciplined workflows, training, and review systems that keep complexity from leaking into production.
Top in AI
Visa and OpenAI bring ChatGPT payments under IPO spotlight
Tether funds NEURA to make robots transact autonomously
White House pauses CAISI public AI reviews amid executive order
Art Directors Guild slams Scorsese for AI storyboarding endorsement
No comments yet. Be the first to share your thoughts!