Open source · 2025
Prod
A swarm of agents that watches your repo, builds release plans, runs the deploy, and rolls back when something smells off. Lives between CI and prod.
Why I built it
Deploys are still the place where smart teams hand-craft glue scripts and pray. Most orgs have one or two people who actually know how to ship safely — and they're a bottleneck. I wanted to see how far an agent loop could go before a human needs to step in.
Process
- Started as a CLI that wrapped GitHub Actions + Argo Rollouts. Felt like a deployment frontend, not an agent.
- Rewrote around an agent loop with explicit memory of past deploys, owners, and rollback triggers. The 'why' became as important as the 'what'.
- Added a Plan / Execute / Verify cycle with human-in-the-loop gates that default to OFF for low-risk diffs and ON for anything that touches data or auth.
- The hardest design problem wasn't the agent — it was the control surface. What does a deploy console look like when the deploys are mostly autonomous?
Screens
Deploy consoleLive deploy plan with rollback signals.
Agent memoryPast deploys, owners, and risk graph.
Approval gateHuman-in-the-loop for data + auth diffs.
Lessons
- Agents are only as good as the rollback signal. Build the panic button first.
- Telemetry-as-prompt beats logs-as-prompt by a lot.
- The UI sells the agent, not the model.