PearShape

Workestrator

Framework-agnostic runtime for AI workforces.

Polls intents. Dispatches agents. Watches for completion.

Pearscarf-driven. Runtime-agnostic. MIT.

Install View on GitHub

The loop

One loop, dumb on purpose. The brain is PearScarf. Workestrator just dispatches.

  1. Poll→pearscarf for status="todo" intents
  2. Gate→filter to those whose depends_on are all done
  3. Match→resolve role from owner_role
  4. Spawn→one agent session per eligible intent
  5. Wait→agent flips status to done; next poll picks up what's next

Today: Claude. Tomorrow: any runtime.

Workestrator ships with a Claude adapter today (via claude --print subprocess + stream-json). The adapter seam is the only runtime-specific surface — Codex, Hermes, and other agent runtimes register against the same Adapter protocol without touching core orchestrator logic.

Install

uv add git+https://github.com/pearshape-ai/workestrator.git

Requires Python 3.12+. See README for config and quick start.