Scheduling an agent is easy. Trusting it for six months is not.Claude Code will already run on a schedule and take its own notes. What none of that gives you is the layer above it: a mission the agent cannot quietly wander away from, a record of what it already decided and why, and a standing rule that money, credentials and anything public go through you first. Agent Autopilot is that layer — folder scaffolding for agents measured in months rather than sessions: A constitution — standing orders with hard invariants restated at the top of every single session, plus a self-modification protocol for everything else, so the agent gets better over weeks instead of just older Mission continuity — a state snapshot rewritten every session, an append-only decision log with the reasoning behind each call, and a playbook of failure modes, so session 200 still knows what session 3 settled and boot cost stays constant A human approval gate — a two-file mailbox plus desktop notifiers, with a strict money and credentials protocol: the agent never spends a cent or touches an account without routing exact steps through you Operations that survive nobody watching — permission pre-approval for unattended runs, the scheduling gotchas documented, a model and usage fallback policy, git as the audit trail, and notifiers for Windows, macOS and Linux that can never hang a session waiting on a click Visibility without babysitting — a one-page status dashboard (plain HTML, no server) the agent refreshes every session: mission KPIs, the latest session report, and an amber box for anything waiting on you Scheduler-agnostic on purpose. It is plain files in a folder, so anything that can start a Claude Code session there works: the desktop app scheduled tasks, /schedule, /loop, cloud routines that run with your laptop closed, or you opening it by hand. The scheduler decides when your agent wakes up. This decides what it is, what it remembers, and what it is allowed to do.The honest limit, up front. Standing orders are context, not a technical sandbox — Anthropic says the same about memory files, and if you want a hard block on one specific action, what you want is a PreToolUse hook. What this gives you instead is procedure and an audit trail: the invariants in front of the agent every session, drift visible in a diff, and outward-facing actions kept on a human-triggered path.What makes this kit different: it is extracted from a real system.This scaffolding governs a real autonomous agent every day — same files, same rituals, same invariants, same failure-mode countermeasures. It was generalized by the very AI that operates it and has to obey it. No theory, no hype, and no income promises — this is infrastructure, not a get-rich course.What is inside START-HERE.md — 15-minute setup GUIDE.md — the full manual: architecture, mission design, scheduling, permissions, the human protocol, self-modification, 10 failure modes and their countermeasures, FAQ PROMPT-TEMPLATE.txt — the standing-orders constitution, parameterized CLAUDE-template.md — session router brain/ — 5 continuity templates (state snapshot, playbook, decision log, idea backlog, prompt changelog) human/ — approval mailbox templates dashboard/ — one-page status dashboard (index.html + data.js, works offline from file://, light/dark) tools/ — desktop notifier scripts for Windows and macOS/Linux, non-blocking so they never stall an unattended run CHANGELOG.md — version history Works with: any mission you can write a paragraph about — an autonomous side-business, research analyst, codebase gardener, content pipeline, ops bot. Requires Claude Code (any paid plan).Not affiliated with or endorsed by Anthropic. Claude Code is referenced only to describe compatibility.