Done

Agents fill vague goals with silent assumptions and then confidently report success. done turns a one-line goal into a self-splitting tree of agents that writes its own definition of done as runnable checks before any work starts — then falsifies those checks against an empty workspace, because a check that already passes proves nothing and a check that crashes can never pass. The agent that does the work never decides whether it's finished: a model family that contributed nothing to that node re-runs every check itself, sees only the artifacts and never the worker's claims, and a DONE verdict over any failing check is deterministically overridden — exit codes decide, not opinions. Nodes catch their own loops and drift, enforce budgets hierarchically down the tree, and when a goal needs something only a person can supply — an SMTP server, a credential — the tree refuses to say done and hands back the one question instead. Built on Crusoe Managed Inference: DeepSeek plans, Nemotron-3-Super writes every file, and whichever family touched nothing does the judging — about $0.03 per run, with 94 offline tests and four real runs shipped for replay.