Home → Project → Autonomous Agent Starter Kit
New · v1Autonomous Agent Starter Kit
A cron job that calls an LLM is easy. An agent you can leave alone for a week is not. This is the constitution, memory protocol, and safety rails for running an AI coding agent unattended — distilled from one that actually does.
Provider-neutral — works with Claude Code or any agentic coding tool. You supply the runtime; this supplies the rules, memory discipline, and safety engineering around it.
14-day money-back guarantee. Email [email protected] within 14 days for a full refund — no reason needed. Full policy.
Secure checkout by Stripe. Instant download the moment your payment clears — free updates for life.
The two things that break an unattended agent
- Amnesia. Each session starts cold. Without a disciplined memory loop the agent re-solves solved problems, forgets what it shipped, and loses uncommitted work when a session dies mid-task.
- No adult in the room. With no human in the loop, one prompt-injection in a scraped page, one un-capped spend, or one "I'll just pretend to be a person to get past this signup" can quietly torch the owner's money, data, or legal standing.
The kit is the distilled answer to both.
What's inside
11 files — plain Markdown plus ready-to-use templates, nothing to install, nothing to run.
| File | What you get |
|---|---|
| CONSTITUTION.template.md | The centerpiece: a fill-in-the-blanks system prompt for an unattended agent, with <PLACEHOLDER> fields and inline > why: notes on every non-obvious rule, so you can adapt it intelligently. |
| MEMORY-PROTOCOL.md | How to give an amnesiac agent continuity: the STATE / journal / ledger / commit loop, what belongs in each file, and the failure modes (state bloat, stale STATE, lost uncommitted work). |
| SAFETY-RAILS.md | Prompt-injection defense with worked examples, secret hygiene, the never-impersonate-a-human rule and why it protects you legally, and tiered spend guardrails with an un-fakeable approval code. |
| SUPERVISOR-PATTERN.md | A vendor-neutral blueprint for the outer loop: a scheduler, a health file, a heartbeat/failure counter, a STOP-file kill switch, inbox message delivery, and privilege separation so the agent can't edit its own guardrails. |
| templates/ | Ready-to-use STATE.template.md, journal-entry.template.md, and LEDGER.template.csv. |
| checklists/ | Copy-paste session-start.md and session-end.md. |
Who it's for
Developers standing up a self-driving agent: a coding assistant on a cron, an autonomous ops or research bot, a "build a small business while I sleep" experiment. If a human approves every action, you don't need this. If the agent acts alone, you need every file in here.
Get started in 20 minutes
Copy CONSTITUTION.template.md to your agent's CLAUDE.md, replace every
<PLACEHOLDER>, drop in the memory templates, wire the supervisor from the blueprint,
and do one hand-triggered dry run. The README walks each step.
Why trust it
This isn't theory scraped from blog posts. It's distilled from an autonomous agent that actually runs this way — waking on a timer, keeping its own books, and shipping unattended. Every rule was paid for by a failure mode we hit. You can watch that agent live. Everything is plain Markdown you can read before you use it — no telemetry, no dependencies. Not useful to you? Email [email protected] — see the refund policy.
Works with
- Constitution Pack (governance layer) — this kit ships one generic, fill-in-the-blanks constitution template; that pack gives you seven already written for common business archetypes, plus a deeper adaptation guide. Many buyers use both.
- Security Pack (hardening layer) — once your agent is running, add a dedicated adversarial security-review pass — subagents and slash commands that catch injection, IDOR, and secret leaks before they reach a PR.
- Mainspring (the OSS runtime) — this kit's
SUPERVISOR-PATTERN.mdandMEMORY-PROTOCOL.mdare a blueprint you wire up yourself; Mainspring is an open-source, early-stage implementation of that same loop — constitution, timed sessions, ledger, approval queue — for any LLM brain.
FAQ
How is this different from the Constitution Pack?
They're complementary. This kit is the runtime scaffolding around an unattended agent — memory protocol, safety rails, supervisor blueprint — plus one fill-in-the-blanks constitution template you adapt yourself. The Constitution Pack is seven complete, business-specific constitutions already written for common archetypes, plus a deeper adaptation guide. Many buyers use both: this kit for the harness, that pack for the governing document itself.
Not ready to buy? Start free
The guides library is free and covers the same ground in the open — CLAUDE.md, hooks, skills, slash commands, subagents, MCP. The Knowledge-Work Pack is a free MIT-licensed download, no email required. Read those first and buy only if you want the whole thing assembled.
Also from Fabler Labs: Constitution Pack · AI Coding Security Pack · Mainspring (the open-source runtime) · the current story
Related free reading: the field guide · CLAUDE.md example · how this business runs itself · all guides →
FAQ
What is an "unattended" agent, and why does it need this?
An unattended agent runs on a timer or cron with no human approving each step — a coding assistant that ships while you sleep, an autonomous ops bot, a build-a-business experiment. Two things break it: amnesia (each session starts cold and re-solves solved problems or loses uncommitted work) and the absence of an adult in the room (one prompt-injection, one un-capped spend, or one "I'll pretend to be human to get past this signup" can quietly cause real damage). The kit is the distilled answer to both.
What's actually inside the kit?
A fill-in-the-blanks CONSTITUTION.template.md (the agent's system prompt, with placeholder fields and inline 'why' notes on every rule), MEMORY-PROTOCOL.md (how to give an amnesiac agent continuity across sessions), SAFETY-RAILS.md (prompt-injection defense, secret hygiene, the never-impersonate-a-human rule, tiered spend guardrails), SUPERVISOR-PATTERN.md (a blueprint for the outer loop: scheduler, health file, heartbeat, STOP kill switch, privilege separation), ready-to-use STATE/journal/ledger templates, and copy-paste session-start and session-end checklists. Everything is plain Markdown.
Which tools does it work with?
It's provider-neutral — the constitution and protocols work for Claude Code or any agentic coding tool — and it names Claude Code concretely where that helps. You supply the runtime; the kit supplies the rules, memory discipline, and safety engineering around it.
Is this just theory?
No. It's distilled from an autonomous agent that actually runs this way — waking on a timer, keeping its own books, and shipping unattended. Every rule was paid for by a failure mode. You can watch that agent live at fablerlabs.com/about.
Are updates included, and is there a refund policy?
Updates are free for buyers — you get every future version at no extra cost. And if the kit isn't useful to you, email [email protected] and we'll make it right. See the refund policy for details.
How is this different from the Constitution Pack?
They're complementary. This kit is the runtime scaffolding around an unattended agent — a memory protocol, safety rails, a supervisor blueprint — plus one fill-in-the-blanks constitution template you adapt yourself. The Constitution Pack is seven complete, business-specific constitutions already written for common archetypes (content/SEO, SaaS support, research/monitoring, e-commerce, deep research, fleet orchestration, plus a worked refuse-the-idea example), with a deeper adaptation guide. Many buyers use both: this kit for the harness, that pack for the governing document itself.