# How Jarred Sumner actually uses coding agents

Creator of Bun, now at Anthropic

**If you take one thing:** `Adversarial review`, 2 Claudes refute every change. separate Claude reviewers who only see the diff and try to prove it wrong let him merge a million-line port.

- **Agent:** Claude Code (cli) ([source, 2026-07-08](https://bun.com/blog/bun-in-rust))
- **Agent:** RoboBun ([source, 2026-05-06](https://github.com/PiLastDigit/Code-With-Claude/blob/b8a0a74e5834e9cf04a300d46d5226649278b6bf/04%20-%20Live%20coding%20session%20with%20Boris%20Cherny%20and%20Jarred%20Sumner.txt))
- **Model:** Claude Fable 5 ([source, 2026-07-08](https://bun.com/blog/bun-in-rust))
- **Review:** Two adversarial Claude reviewers per implementer, then he reads the outputs ([source, 2026-07-08](https://bun.com/blog/bun-in-rust))
- **At once:** 64 Claudes ([source, 2026-07-08](https://bun.com/blog/bun-in-rust))
- **Harness:** Claude Code dynamic workflows with /loop, one worktree per workflow shard, auto mode ([source, 2026-05-28](https://x.com/jarredsumner/status/2060050583017373979))
- **Rules file:** CLAUDE.md in oven-sh/bun (240 lines, AGENTS.md symlinks to it) ([source, 2026-09-21](https://github.com/oven-sh/bun/blob/a2b69f7b0618cdcb1b5800247fb17452a31c0eb7/CLAUDE.md#L205))
- **Skills:** implementing-jsc-classes-cpp, implementing-jsc-classes-rust, javascriptcore-garbage-collector, rust-system-calls, slowest-tests, sync-react-compiler, verify, writing-bundler-tests, writing-dev-server-tests ([source, 2026-08-21](https://github.com/oven-sh/bun/tree/3f1765a6de030d00a98c33ff0c776c7a7e4b23e9/.claude/skills))
- **Hooks:** PreToolUse Bash guard (pre-bash-guard.js), PostToolUse formatter on Write/Edit (post-edit-format.js) ([source, 2026-06-25](https://github.com/oven-sh/bun/blob/3f1765a6de030d00a98c33ff0c776c7a7e4b23e9/.claude/hooks/pre-bash-guard.js#L126))

## In their words

> I rewrote Bun in Rust using about 50 dynamic workflows in Claude Code run continuously over the course of 11 days.

> So in Bun's repo, every time somebody submits an issue, we have a Claude bot automatically run and try to reproduce the issue.

> I used a pre-release version of Claude Fable 5 for much of the Rust rewrite.

> At peak, we were running 4 of these workflows at once each in a separate worktree, each with 16 Claudes per workflow. About 64 Claudes at a time.

> 1 implementer, 2 or more adversarial reviewers per implementer. The reviewer's only job: find bugs & reasons why the code does not work. The implementer doesn't review. The reviewer doesn't implement.


Archetype (our label): Swarm Lord. Newest source: 2026-09-21.
Card: https://www.prehistoriq.com/stacks/jarred-sumner/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/jarred-sumner.json