swarm Install

One goal in.
A verified result out.

swarm runs a fleet of sub-agents across your codebase and stops only when the work is proven.

Install GitHub
goal sub-agents proven

A run

Six phases, and it will not leave early.

  1. 01

    Read the world

    Your repo, your tools, what the last run learned.

  2. 02

    Map it

    One agent per area, then a graph of the work.

  3. 03

    Find the cause

    The exact failing path, with the evidence.

  4. 04

    Break it

    A second agent tries to prove that wrong.

  5. 05

    Fix in isolation

    Every fixer in its own worktree and branch.

  6. 06

    Sweep again

    Checks green, then hunt for what is left.

The strongest thing you can do is give an agent a way to review its work.

That is what swarm is.

So this one argues
with itself.

No answer counts until a second agent, fresh context, has tried to kill it.

root cause

The connection pool is exhausted under load.

the second agent

Pool raised to 40. The timeout persists.

Refuted

actual cause

The retry fires before the write it is retrying commits.

Built for runs
that last hours.

Isolated

Parallel fixers never share a working tree. Each gets its own worktree and branch.

Guarded

Never force-push, never push to main, never merge. Anything that costs money stops and asks.

Stacked

Dependent work opens as a stack of pull requests, in merge order.

Durable

A ledger on disk, so a restart resumes instead of redoing finished work.

Clone it into your skills folder.

# once, for Claude Code or Cursor
git clone https://github.com/SaarthurR/swarm.git ~/skills/swarm
ln -s ~/skills/swarm ~/.claude/skills/swarm

Then hand it a goal.

swarm: checkout drops orders under load. find out why and fix it properly.