Agent Atlas

The Builders

Boris Cherny

Created Claude Code at Anthropic.

Run several sessions at once, plan before building, and give the agent a way to check its own work.

Steal this

  • Start most sessions in plan mode.
  • Every mistake becomes a line in the shared CLAUDE.md.
  • Pre-allow safe commands instead of turning checks off.
  • Give the agent a way to verify its own work.

The lesson and the takeaways are our summary. Everything below is theirs, with where they said it.

The setup

Factual
Machines

A terminal with five numbered tabs, plus the web and his phone.

Software

Claude Code, mostly as it ships. Shared slash commands, subagents and hooks checked into git.

Models

Opus 4.5 with thinking for everything, at that date.

As of January 2, 2026. Setups change; follow the links.

Their gear

Links go to the maker's own page. We earn nothing from them. Every machine is on the gear shelf.

How they work

Factual
  1. Five sessions in the terminal, with system notifications when one needs him. (Boris Cherny on X)
  2. More sessions on the web, some started from his phone in the morning. (Boris Cherny on X)
  3. Most sessions start in plan mode. (Boris Cherny on X)
  4. One shared CLAUDE.md in git. Every mistake becomes a line in it. (Boris Cherny on X)
  5. Repeated workflows become slash commands and subagents. (Boris Cherny on X)
  6. Safe commands are pre-allowed instead of turning permission checks off. (Boris Cherny on X)
  7. The most important habit is letting the agent verify its own work. (Boris Cherny on X)
  8. By September 2026 he ran Opus 5.5 on medium effort for most work, low for speed and xhigh for more thinking. (Boris Cherny on X)
  9. In October 2026 he kept medium about 95% of the time and raised Opus to xhigh or max for complex visual work. (Boris Cherny on X)

Where they talk and write

Interviews 19

13 older

Writing 19

13 older

Each link goes to the original page, newest first.

Habits worth copying