# Agents Never Sleep (ANS)

> Open-source harness + Agent Skill that lets a coding agent run a ticket backlog to
> completion UNATTENDED: a durable per-ticket state machine, an ASK/PARK/HALT autonomy
> contract (decisions a machine shouldn't make alone get parked for the human), deterministic
> test gates, git-backed reversibility, a heartbeat watchdog, and a ranked morning report.
> MIT-licensed, Python stdlib only. Built by Tokonomix.ai (InterIP Networks).

## What it is
- A Python package (`agents_never_sleep`) + portable Agent Skill (SKILL.md) — install:
  `pip install git+https://github.com/TokonoMix/agents-never-sleep@v1.0.0`
- The agent works tickets one by one; each ticket only counts as done when its deterministic
  test gate passes. Failing work is reverted, classified and reported — never hidden.
- ASK is forbidden while unattended: a blocking question becomes a PARK (recorded, surfaced
  in the morning report) and the run moves on. HALT stops the whole run when continuing
  could do harm.

## Honest limits (stated deliberately)
- It does not guarantee correct code: gates catch what tests catch; a human reviews the
  morning report.
- Permission-enforcement hooks exist for Claude Code only today. Hermes/Aider adapters are
  roadmap items — not claimed until they ship.
- No throughput numbers are published: benchmarks are published as reproducible methodology
  or not at all.
- The watchdog/gates/revert paths REDUCE the ways an overnight run can go wrong; they do not
  make autonomy risk-free.

## Managed ANS (in development — not yet available)
A governance layer for teams running fleets of unattended agents, via the Tokonomix gateway
(https://tokonomix.ai/): one token per fleet (no raw provider keys in agent configs), a hard
monthly budget cap enforced at the gateway, central model/provider allow-lists, EU
data-residency routing, central billing and usage visibility. The open-source harness stays
fully functional on your own keys — the paid layer is governance, not gated features.

## The family (supersystem)
ANS is the overnight-execution layer of a family of independent agent tools by the same maker
(each stands alone on your own keys; a Tokonomix key ties them together):
- Tokonomix — cross-vendor LLM consensus + model routing, the paid hub: https://tokonomix.ai/
- AgentsWeaver — cross-machine, signed agent mesh (scale-up): https://agentsweaver.ai/
- agentixmesh — local single-machine agent trust layer: https://agentixmesh.ai/
- DCP — open Development Coordination Protocol: https://devcopro.org/
- ZelixAI — embeddable AI chat/RAG for websites: https://zelixai.ai/
- Gallery — AI image and creation studio (no public site yet)
Details: https://agentsneversleep.ai/supersystem/

## Pages
EN is canonical; each page is also published in nl, de, fr, es, tr (same path, locale prefix).
- /en/ — home: the mission, the six mechanisms, the ASK/PARK/HALT contract, what it doesn't do
- /en/how-it-works/ — deep walkthrough of one run: launch preflight, per-ticket state machine,
  deterministic gates, ASK/PARK/HALT decision points, the watchdog, drain, the morning report
- /en/faq/ — 12 straight-answer questions: safety, supported agents, licensing, Managed ANS,
  what gets parked, whether it can push to production
- /en/managed/ — Managed ANS status page (in development, not yet available): the five planned
  fleet-governance capabilities and the honest OSS-vs-managed split
- /en/supersystem/ — how ANS fits the wider InterIP agent family (Tokonomix, AgentsWeaver,
  agentixmesh, DCP, ZelixAI)

## Key links
- Source + README: https://github.com/TokonoMix/agents-never-sleep
- License (MIT): https://github.com/TokonoMix/agents-never-sleep/blob/main/LICENSE
- Tokonomix (multi-model consensus gateway, same maker): https://tokonomix.ai/