Frequently asked questions

Agents Never Sleep (ANS) is an open-source governance layer that lets a coding agent run a backlog of tickets to completion unattended. It gives the run a durable per-ticket state machine, deterministic test gates, git-backed reversibility, and an ASK/PARK/HALT autonomy contract, so a single unanswerable question never freezes the rest of the backlog. It does not write code itself — it governs how an agent that writes code behaves while nobody is watching.

ANS reduces the risk, it does not eliminate it. Deterministic gates catch what your tests catch, git snapshots make every change reversible, and deny-hooks block outright-irreversible actions like force-pushes or destructive SQL. But the deterministic gate is only as good as your test suite, and a wrongly-classified assumption is still possible — which is exactly why every assumption the run makes is committed so it can be reverted in daylight. Your execution environment (a container, a throwaway checkout, a least-privilege user) is still the primary safety layer; ANS assumes you have one.

Claude Code is the only platform where the enforcement hooks (deny-irreversible, never-stop, never-ASK) are live-verified on the real tool. Gemini CLI, OpenAI Codex CLI, GitHub Copilot CLI, Cursor and Windsurf are built to each platform's documented hook contract and covered by the hermetic test suite, but not yet confirmed on the real tool. Aider is supported through a launcher-preset wrapper (launch flags plus git-revert reversibility) rather than native hooks. Hermes support is on the roadmap, not shipped.

The harness classifies the failure. If it looks introduced by the current diff, the change is reverted to the last green commit and the ticket is recorded FAILED_RETRYABLE so it can be attempted again. If the failure looks pre-existing, flaky, or environmental, the work is kept but confidence is downgraded and it's noted as a blind spot rather than a clean pass. ANS never deletes or skips a failing test to force a green result — that would hide exactly the signal the gate exists to surface.

Anything with a large blast radius: which direction a database migration should go, a change to a public or shared API contract, anything touching a security or tenant-isolation boundary, and anything involving money, billing, or pricing. A park is not a failure — it records why, the candidate interpretations, and the exact decision waiting for a human, then the run moves on to the next independent ticket.

A single ranked summary (night-report.md) written when the run ends. It lists what's done and trusted, what's done but needs daylight review, what's parked (with the exact decision needed), what's blocked by the environment, and any blind spots the run hit along the way. It exists so a human reviews decisions and outcomes, not a raw stream of logs.

No. The open-source harness is MIT-licensed, Python standard library only, with zero runtime dependencies, and it runs on whatever credentials your coding agent already uses. There is an optional delegated review step that can call the Tokonomix gateway for a second opinion on a high-risk diff, but that's opt-in and budget-gated — the core harness works fully on your own keys without it.

MIT. The source is public on GitHub.

A bare loop has no principled answer for what to do when the agent isn't sure — it either stops and waits for nobody, or it guesses and might do something irreversible. ANS replaces that gap with a durable per-ticket state machine, a deterministic gate as the only hard block, git-backed snapshot-and-revert on every change, and an explicit ASK/PARK/HALT contract enforced at the tool layer rather than left to the agent's judgment mid-run. The loop is the same two commands; what's different is everything that governs what happens between them.

Managed ANS is in development, not yet available. It's planned as a governance layer on top of the open-source harness for teams running fleets of unattended agents: one token per fleet, a hard budget cap enforced at the gateway, model and provider allow-lists, EU data-residency routing, and central billing and usage visibility. The open-source harness stays fully functional on your own keys whether or not you ever use the managed tier. Read more →

No — by design. ANS never pushes to a remote branch; work stays as local, reversible git commits for a human to review and push deliberately. Deny-hooks additionally block outward or irreversible actions like force-pushes and remote branch deletion at the tool layer. A human is always the one who decides when work leaves the machine.

Open an issue on the GitHub repository. That's also where roadmap items like broader platform enforcement and the PyPI publish are tracked.