Boomerang — the agent that calls you back

Coding agents now run 10–60 minutes unattended, but supervision is still synchronous: a terminal you babysit defeats the point, and a chat window you poll is worse. Every "autonomous" run ends the same way — you checking whether it's done. Boomerang inverts the flow. It's a two-way voice bridge built on the AssemblyAI Voice Agent API. Brief a coding task by voice, then walk away. A headless worker does the fix on the real repo — clone, reproduce, patch, test, push. When it finishes — or hits a wall — the agent initiates a voice callback: a chime and a 15-second spoken summary. It holds the line for your verbal go/no-go, logged as a consent receipt (spoken utterance + transcript hash + timestamp) before any side effect runs. Say "ship it" and a real PR opens. Under the hood: two orchestrated voice sessions (brief + agent-initiated callback), keyterms for repo jargon, JSON-schema tool.call as the consent gate, and turn detection tuned for one-word approvals. Failure is a first-class path — the callback tells you what broke and offers retry or cancel. Voice becomes an interrupt channel, not a chat window — the same insight that made pagers beat dashboards for ops, applied to agent supervision. Every action traces to the exact spoken words that authorized it.