Blackout

Blackout lets AI agents act on their own without doing anything twice. Agents are starting to send emails, issue refunds and change systems unattended; when one crashes mid-job it re-does work it already did, and it says "done" when it isn't. Blackout sits between the agent and the real world: it writes every side effect down with a stable ID before it happens, checks the outside system's receipts after a crash to see what really happened, and lets only plain code (never the model) decide the job is done. When it can't tell, it escalates to a human. We built it two ways today. The full runtime: a Nemotron coordinator writes a definition of done that Blackout freezes, a Nemotron repair worker diagnoses and fixes a broken service, we SIGKILL it mid-delivery, and before its replacement continues, Blackout checks each unfinished action against the service's receipts. In 10 paired live runs on a service that can't report job status, Blackout finished cleanly 10/10; the same agent without it 0/10 (16 duplicate deliveries). And a drop-in MCP gateway, published on npm as blackout-mcp, that any MCP agent (including Claude Code) can put in front of its tools. Measured: a Nemotron agent told not to send duplicates, restarted mid-job, duplicated emails in 6 of 10 runs because it re-invented its idempotency key each time; through Blackout, 0 of 10. All agents run on NVIDIA Nemotron 3.5 Lightning via Crusoe Managed Inference, about half a cent per mission. Site with a 3D explainer and recorded live race: https://blackout.black. Everything in the repo was built during the event.