// the take
Multi-agent orchestration framework that spins up a temporary coordinator to split a task across registered 'specialist' agents using the A2A agent-card protocol, meant to run inside Claude Code, Codex, Cursor and similar hosts. Aimed at developers building multi-agent workflows who want task decomposition and specialist routing without writing their own orchestrator.
Uses the A2A v1.0 agent-card standard for identity/capability checks instead of a proprietary agent format, which is a real interoperability choice. Pins a chosen agent release by digest so a silent upstream change doesn't alter behavior mid-run. Backend-agnostic across several hosts and model providers rather than locking you into one CLI.
The README is almost entirely marketing narrative built around unverifiable headline numbers (token savings, match-rate percentages) with no linked benchmark or reproducible methodology — none of it should be taken at face value. The installer is explicitly written to be read and executed by an AI agent, carries a hidden instruction block at the top of the file, writes config into nearly a dozen different tool directories, and auto-answers a security trust prompt on the user's behalf — a genuinely bad pattern independent of intent. There's no actual orchestration logic visible in the README or top-level tree; what's there is mostly skill/workflow markdown and vendored third-party packages, not auditable application code. The real engine sits behind a commercial Desktop app, Hub and Cloud product, with this repo functioning mainly as the open distribution and install mechanism rather than a self-contained open-source tool.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →