Swarm breaks a spec into tasks, runs them in parallel as autonomous coding agents, reviews the output, and opens pull requests. You provide a spec (Markdown) and credentials; the swarm handles planning → execution → review → PR. It routes across providers (Anthropic, OpenAI, Gemini, Bedrock, and more) per phase and per task.

Pipeline order: plan → execute → [test] → [review + fix] → integration validation → [PR].

1. 1
   Make sure you belong to an organization at /orgs. Organizations are provisioned centrally by a platform admin; once you belong to one, all credentials, repos, and runs are shared across its members.
2. 2
   Open /settings and fill in:

   • GitHub token — needs repo and workflow scopes
   • A provider key — e.g. an Anthropic API key (or Bedrock / OpenAI / Gemini credentials)
   • Repos — one or more GitHub clone URLs
3. 3
   Go to /new, paste a Markdown spec, pick options, click Start run.
4. 4
   Watch the run on its detail page — tasks stream live, each with its own logs, diff, and cost.
5. 5
   When done, the engine opens a PR (unless you checked Skip PR creation).