// the find
affaan-m/claude-swarm
Multi-agent orchestration for Claude Code — decompose tasks, coordinate agents, visualize everything in a rich terminal UI
claude-swarm wraps the Claude Agent SDK to run multi-agent task execution with dependency-aware scheduling and a live terminal dashboard. You describe a task, Opus decomposes it into a DAG, Haiku agents execute subtasks in parallel waves, then Opus reviews the combined output. Built for a hackathon, but the architecture is sensible.
The model selection strategy is smart — Opus for planning and QA, Haiku for worker execution. The file-locking mechanism is the right answer to the obvious agent-collision problem. Session recording as JSONL events with a replay command is genuinely useful for debugging what went wrong. The dependency graph approach (topological sort via NetworkX) is cleaner than prompt-chaining or a round-robin queue.
This is a hackathon project from February 2026 that hasn't been touched since — 'last push: 2026-02-11' on a tool that wraps an SDK still in 0.x means you're buying someone's weekend experiment built on an unstable API. The quality gate is Opus reviewing text output from Haiku agents, not running the code or tests, so it catches consistency issues but not correctness ones. Pessimistic file locking will serialize your whole workflow the moment any two tasks touch shared files, and there's no mechanism to break a task apart to avoid that. 44 tests for a concurrent multi-agent system is thin — the interesting failure modes (budget race conditions, partial wave failures, orphaned file locks) are probably not covered.