Use cases

Before touching a large codebase, spawn three agents: one to map the current structure, one to find the highest-risk areas, one to draft a refactor plan. The three agents can message each other — the planner asks the explorer for clarification, the risk-finder flags concerns. You get a synthesis in minutes, not hours.

A new feature spec is ready. Spawn a frontend-developer, a security-auditor, and a code-reviewer agent. Each reviews the spec from their angle. They debate via P2P questions, converge on concerns, and surface a unified review document. No more scheduling three meetings.

A CI run failed across 47 tests. Spawn an agent to categorize the failures (regression, flaky, infrastructure, real bug), then attach a debugging agent to the real-bug subset. Each debugging agent works in parallel and posts its findings to the orchestrator.

Spawn a deep-research fleet that runs for hours, not minutes. Per-fleet timeouts mean individual agents can think deeply without breaking the wall. Event log gives you a complete audit trail of who did what. Pick up where you left off after a restart.

Migrating from one library to another. Spawn an agent to do the migration, another to write tests, another to verify no regressions. The migrator and the verifier can message each other: 'I changed this file, does the test still pass?'

A long-running fleet stays in the background. If an alert fires, the alert handler broadcasts to the fleet. Peers with relevant skills (database, network, auth) respond. If a peer can fix it, they reply with the fix. If not, they escalate to a human via request_help.