AnvilWorks is a Python daemon that orchestrates AI coding agents on your behalf. It plans, executes, retries, and escalates — so you can hand it intent and walk away.
Loop = trigger + goal + body + brakes. A self-firing, self-verifying, self-limiting unit of work that runs coding agents, entire projects, or Armature multi-agent teams.
Heavy use of AI coding tools today looks like 4–5 terminal tabs, each babysitting a separate agent on a different project. AnvilWorks replaces that with one supervised execution daemon: automatic sequencing, live visibility, git integration, failure triage, and a persistent inbox for the decisions that actually need you.
A loop never stops because the agent claimed it was done. It stops when an independent check passes, when it hits a hard limit, or when it escalates a concrete question to you.
The body is pluggable: a coding agent for code work, an AnvilWorks project for phased delivery, or an Armature multi-agent team for heavy multi-agent deliberation. AnvilWorks wraps each body in scheduling, verification, brakes, and escalation.
Every loop must carry brakes. A loop that can't reach its goal is guaranteed to stop and escalate — never spin forever.
AnvilWorks is a Python 3.11+ project. Install it editable with pipx or pip, start the daemon, and create your first project or loop.
The repository is public. Use it, fork it, open issues, or adapt the harness registry and loop engine for your own agents.
github.com/bryansparks/anvilworks