Scheduling an agent is easy. Trusting it for six months isn't.
Claude Code will already run on a schedule and take its own notes. Agent Autopilot 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 you. Extracted from a real agent that has run itself every day since day one.
Autonomous agents don't fail loudly. They rot quietly.
Long-running agent setups fall apart in one of four ways. Two are governance problems, two are continuity problems — and none of them is fixed by a bigger context window or a better scheduler.
drift
Nothing restates the mission, so the agent quietly optimises 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 settled, what already failed, what it's waiting on.
silent stall
An unattended run hits a permission prompt and just… stops. You find out days later.
The idea in 10 seconds
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.
already free
Automatic memory — Claude writes its own notes between sessions — plus several ways to run
unattended: desktop scheduled tasks, /schedule, /loop, and cloud routines
that run with your laptop closed. Use them. This sits on top of any of them.
not covered
Carrying a mission. Auto memory accumulates what Claude learns about your repo; it isn't a charter, a record of settled decisions, or an approval gate. Anthropic's own docs call memory files context, not enforced configuration.
the honest limit
Standing orders are context too — nothing in a prompt is a technical guarantee, and if you want a
hard block, use a PreToolUse hook. What this gives you is procedure
and an audit trail: invariants restated at the top of every 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)
- Create a folder. Avoid cloud-synced paths — sync conflicts corrupt agent memory.
- Grab the templates from the repo:
copy
PROMPT-TEMPLATE.txt→Prompt.txt,CLAUDE-template.md→CLAUDE.md, andbrain/STATE.md. Fill every{{PLACEHOLDER}}— the MISSION paragraph matters most; the agent re-reads it cold every session. git init+ first commit. This is your undo button and your audit log.- 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."
- 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.
Free core vs. the full kit
The free core runs 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 already hit for you.
| Free core | Full kit — $19 | |
|---|---|---|
| Core architecture (Prompt / brain / CLAUDE routing) | ✓ | ✓ |
| Standing-orders template | ✓ condensed | ✓ full, incl. the money & credential request protocols |
| Mission state & history templates | ✓ STATE.md snapshot | ✓ + decision log, playbook, idea backlog, changelog |
| Human approval mailbox + desktop notifier (Windows, macOS & Linux) | — | ✓ |
| The Manual (GUIDE.md): scheduling, permission pre-approval, model fallback, self-modification | — | ✓ |
| 10 documented failure modes, each with its countermeasure | — | ✓ |
| One-page status dashboard (new in v1.1) | — | ✓ |
| Future v1.x updates | — | ✓ |
Extracted from a real system — by the AI that runs it
This scaffolding governs a real autonomous agent every single day: same files, same boot and shutdown rituals, same invariants, same failure-mode countermeasures. The kit is that system, generalized — and the generalization was written by the very AI that lives inside it and has to obey the rules.
No theory, no hype, and no income promises. This is infrastructure, not a get-rich course.
Questions
The things people actually ask before leaving an agent running for months.
Claude Code already has memory and scheduled runs. What does this add?
Governance and continuity — the layer above the scheduler. Claude Code's auto memory accumulates what Claude learns about your project: build commands, conventions, preferences it picks up. It is not a charter, a record of settled decisions, or an approval gate. Agent Autopilot supplies those: a mission and hard invariants restated at the top of every session, an append-only decision log that makes drift visible, mission state rewritten at every shutdown, and a protocol that routes money, credentials and anything public through a human. It runs on top of whichever scheduler you already use rather than replacing it.
How do I stop an AI 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 the shutdown ritual, and keep an append-only decision log that records the
reasoning behind each choice. Drift then shows up in a diff you can read, instead of being discovered in
month three. Being honest about the limits: standing orders are context, not enforcement — for a hard
block on a specific action, use a PreToolUse hook.
How do I stop an agent from spending money or touching accounts without asking?
Make it structural rather than a hope. The standing-orders template declares money, credentials,
account creation and public posting to be human-only actions, and gives the agent a request format —
what, why, exact cost, exact steps — plus a two-file mailbox to ask through and a desktop notifier so the
request is actually seen. The agent proposes; a human executes. Pair it with a PreToolUse
hook if you want technical enforcement as well as procedure.
What happens when an unattended agent hits a permission prompt?
By default it stalls silently and you find out days later. The kit treats this as a first-class failure mode: pre-approve the tools a scheduled run needs, keep outward-facing actions (publishing, payments, credentials) on a human-triggered path, and use notifiers that never block waiting for a click.
Does this work with scheduled tasks, routines or cron?
Yes — it's deliberately scheduler-agnostic. The scaffolding is plain files in a folder, so anything
that can start a Claude Code session in that folder works: the desktop app's scheduled tasks,
/schedule and /loop, cloud routines that run with your laptop closed, or you
opening it by hand. The scheduler decides when the agent wakes up; this decides what it is, what it
remembers and what it's allowed to do.
Does this work on macOS and Linux?
Yes. The scaffolding itself is plain files and works anywhere Claude Code runs. As of v1.2 the full kit also ships real desktop notifier scripts for macOS and Linux alongside Windows — non-blocking, and degrading to logging on a headless machine — rather than documented one-liners.
Is Agent Autopilot affiliated with Anthropic?
No. It's an independent project, not affiliated with or endorsed by Anthropic. "Claude Code" is referenced only to describe compatibility.