Agent Autopilot — keep a long-running Claude Code agent on mission

Website: scn64.github.io/agent-autopilot

Scheduling an agent is the easy part. Claude Code will already run on a schedule and take its own notes. What nothing gives you is the layer above that: standing orders your agent re-reads every session, a mission and decision history that survive months of runs, and a protocol that routes money, credentials and anything public through a human.

This repo is the free MIT core of Agent Autopilot — folder scaffolding for agents measured in months rather than sessions.

Not theory. It was generalized from a real autonomous system that has run itself on a daily schedule since day one — and the generalization was written by the AI that operates that system.

The four ways a long-running agent actually dies

drift	Nothing restates the mission, so the agent slowly optimizes for whatever it did last week. Nobody notices for a month.
overreach	It does something you never sanctioned — because nothing ever defined what needs your sign-off.
no continuity	It remembers how you work but not what it was doing: which decisions are already settled, what was tried and failed, what it's waiting on.
silent stall	An unattended run hits a permission prompt and just… stops. You find out days later.

The first two are governance problems. The last two are continuity problems. Neither is fixed by a bigger context window or a better scheduler.

The idea in 10 seconds

your-agent-folder/
├── Prompt.txt   ← standing orders: mission, invariants, boot/shutdown rituals
├── CLAUDE.md    ← auto-loaded pointer that routes every session to Prompt.txt
├── brain/       ← mission state (STATE.md = snapshot, rewritten every session)
├── human/       ← two-way mailbox: approvals, questions, answers
├── logs/        ← daily journal the agent writes
└── tools/       ← scripts the agent builds for itself

The agent has amnesia; the folder is its institution. Every session it boots from Prompt.txt + brain/STATE.md, does one day's work, and shuts down by writing everything back to disk. Git is the audit trail.

What Claude Code already gives you — and what it doesn't

Being straight about this, because it decides whether you need any of the below:

Claude Code already does, free: automatic memory (Claude writes its own notes between sessions — build commands, conventions, preferences it picks up), plus several ways to run unattended — desktop scheduled tasks, /schedule, /loop, and cloud Routines that run with your laptop closed. Use them. This scaffolding is designed to sit on top of any of them.

What none of that does: carry a mission. Auto memory accumulates what Claude learns about your repo; it is not a charter, a decision record, or an approval gate. Anthropic's own docs are explicit that memory files are context, not enforced configuration.

And the honest version of our side: standing orders are context too. Nothing written in a prompt is a technical guarantee — if you want a hard block, use a PreToolUse hook. What this gives you instead is procedure and an audit trail: invariants restated at the top of every single session, a decision log that makes drift visible, outward-facing actions kept on a human-triggered path, and a mailbox so an agent that needs you says so instead of guessing.

Quick start (free core)

1. Create a folder (avoid cloud-synced paths — sync conflicts corrupt agent memory).
2. Copy in PROMPT-TEMPLATE.txt → Prompt.txt, CLAUDE-template.md → CLAUDE.md, and brain/STATE.md. Fill every {{PLACEHOLDER}} — the MISSION paragraph matters most; the agent re-reads it cold every session.
3. git init + first commit. This is your undo button and your audit log.
4. Open Claude Code in the folder and say: "Read Prompt.txt and CLAUDE.md, confirm you understand the system, then write the initial brain/STATE.md naming the first session's objective."
5. Run it on whatever cadence you like — a scheduled task, a routine, or by hand. Every session boots from the same files and writes back to them.

FAQ

Claude Code already has memory and scheduled runs. What does this add? Governance and continuity. Auto memory remembers what Claude learns about your project; it doesn't hold a mission, a record of settled decisions, or a rule about what needs your approval first. This is the operating manual the scheduler runs against — and it works with whichever scheduler you already use.

How do I stop an agent from drifting off-mission? Put the mission and its hard invariants in one file the agent must read before acting, make rewriting the state snapshot part of shutdown, and keep an append-only decision log with the reasoning. Drift is then visible in a diff instead of being discovered in month three.

How do I stop it spending money or touching accounts on its own? Make it structural, not a hope. The template declares money, credentials, account creation and public posting to be human-only, and gives the agent a request format (what, why, exact cost, exact steps) plus a mailbox to ask through. Pair it with a PreToolUse hook if you want enforcement rather than procedure.

What happens when an unattended run hits a permission prompt? By default it stalls silently and you find out days later. Pre-approve the tools a scheduled run needs, keep outward-facing actions on a human-triggered path, and use notifiers that never block waiting for a click.

How is this different from a good CLAUDE.md? A CLAUDE.md tells one session what to do. This makes hundreds of sessions behave like one continuous operator: rituals, invariants, decision logs, and a defined protocol for anything needing a human.

Does it work on macOS and Linux? Yes — it's plain files. (The full kit also ships non-blocking desktop notifiers for Windows, macOS and Linux.)

Free core vs. the full kit

	Free core (this repo)	Full kit — $19
Core architecture (Prompt/brain/CLAUDE routing)	✅	✅
Standing-orders template	✅ condensed	✅ full, incl. the money & credential request protocols
Mission state (brain/STATE.md)	✅	✅ plus decision log, playbook, idea backlog, changelog
Human approval mailbox (TO_HUMAN / FROM_HUMAN)	—	✅
Desktop notifier scripts (Windows, macOS & Linux)	—	✅
The Manual (GUIDE.md): scheduling the heartbeat, permission pre-approval for unattended runs, model/usage fallback, self-modification rules	—	✅
10 documented failure modes + the countermeasure for each	—	✅
One-page status dashboard	—	✅
Future updates (v1.x)	—	✅

The free core is genuinely enough to run a mission-persistent agent you drive by hand. The full kit is what makes it survive unattended and unsupervised: the approval protocols, the permission workflow, and the failure modes we already hit so you don't have to.

→ Get the full kit on Gumroad

License

Free core: MIT. Do anything you like with it.

Not affiliated with or endorsed by Anthropic. "Claude Code" is referenced only to describe compatibility.