What it covers
- Controller and plan — A deterministic orchestrator decides which agent runs, with what context and what it may do; the model never picks its own permissions.
- Tool contracts — JSON-schema inputs and outputs for every tool, validated on both sides, with rate limits and per-tenant scopes.
- Human-in-the-loop — Approval queues for irreversible actions, with the proposed action, the evidence and a one-click accept or reject.
- Durable runs — Steps checkpointed so a job can resume after a crash, a deploy or a rate-limit pause without repeating side effects.
How we work
We map the workflow on a whiteboard first — which steps are deterministic, which need a model, which need a person — and only then write prompts.
