Your agent works the backlog.
You read the morning report.
Agents Never Sleep is an open-source harness + Agent Skill that lets a coding agent run a ticket backlog to completion unattended — with a durable per-ticket state machine, deterministic test gates, git-backed reversibility and an autonomy contract that parks what a machine shouldn't decide alone.
Source on GitHub MIT · Python stdlib only · v1.0.0
5 completed (each git-snapshotted, gates green), 1 parked — the schema-direction call it wouldn't make without you.
Mechanism, not magic
Six things the harness actually does
Every capability below is a concrete mechanism in the codebase — named here by the thing that implements it. The full walkthrough lives on How it works.
Durable ticket state
Each ticket moves through an explicit state machine that survives crashes and restarts. A run resumes where it stopped — it doesn't start over.
ASK / PARK / HALT
Decisions a machine shouldn't make alone — money, schema direction, security boundaries — are parked for the human, and the run moves on to the next ticket.
Tests decide, not vibes
A ticket only counts as done when its test gate passes. A failing gate reverts the work; the failure is classified and reported, not hidden.
Everything reversible
Work is snapshotted per ticket. Anything the night produced can be reverted commit-by-commit in the morning.
Survives frozen runs
A watchdog wraps the run by default and restarts it resumably when the agent freezes mid-run — for example during a provider overload wave.
Morning report
The run ends with a ranked report: what landed, what parked and why, what needs your judgement. You review decisions, not diffs-at-random.
The autonomy contract
Never block the run on a single ticket
Unattended doesn't mean unaccountable. The contract defines exactly what happens when the agent hits something it shouldn't decide alone.
Forbidden while unattended. The harness never sits waiting on a question at 3 a.m. — a question that would block becomes a park instead.
Record and move on. Money, schema direction, security boundaries, public API contracts: parked with the reasoning, surfaced in the morning report for a human decision.
Stop the whole run. Reserved for conditions where continuing could do harm — the run ends, states are preserved, the report says why.
What it doesn't do
The honest part
- It doesn't guarantee correct code. Gates catch what tests catch. The morning report exists precisely because a human still reviews the outcome.
- Enforcement hooks are Claude Code-only today. The harness and skill are portable; the hard permission-enforcement layer currently exists for Claude Code. Hermes and Aider adapters are on the roadmap — until they ship, we won't claim them.
- No throughput numbers on this page. We publish benchmarks as reproducible methodology or not at all. A controlled agent-vs-harness comparison is planned; its numbers will appear when they exist.
- Reduced, not eliminated. The watchdog, gates and revert paths reduce the ways an overnight run can go wrong. They don't make autonomy risk-free — the contract is designed for exactly that reality.
Get started
Two minutes to a first supervised run
- Install the package and the Agent Skill (both live in the same repo — see the README).
- Write a small backlog: a directory of self-contained tickets with a deterministic gate each.
- Run supervised first. Watch one batch land with you present; go unattended when the gates and parks behave the way you expect.
Managed ANS
The open-source harness is complete and free, on your own API keys — that doesn't change. Managed ANS is the governance layer we're building on top for teams running fleets of unattended agents, via the Tokonomix gateway: one token per fleet, a hard budget cap, model allow-lists, EU data-residency routing and central billing. Read what's being built →