swarm

What problem does it solve? A single search pass over a codebase, document set, or research area has blind spots and quits early, so rare bugs, edge cases, and prior art get missed. Swarm fans out many independent finder subagents across distinct slices or angles, then dedups and unions their findings until no new results appear. ## Core Features & Use Cases - Parallel blind finders: Dispatches batches of subagents in a single message, each searching a different slice (module, file batch) or angle (data-flow, error-handling, spec-conformance) with no visibility into each other. - Loop-until-dry convergence: Runs rounds of searches, dedups every find against a running seen-set, and stops only after K consecutive rounds produce zero fresh findings. - Structured findings and coverage reporting: Forces a uniform find schema (location, what, severity, evidence) and outputs a severity-sorted deduped table plus a convergence curve and an explicit list of what was not searched. - Use Case: Ask for an exhaustive security audit of src/auth — swarm launches finders sliced by data-flow, error paths, and spec deviation, loops until two dry rounds, and returns every deduped vulnerability with evidence. ## Quick Start Ask the agent to swarm this codebase and find every security bug, running rounds of parallel finders until no new findings appear.