The field guide · Starter kit

The agentic OS, as real files

Everything the nine-layer guide describes — a checkable constitution, separation of powers, a daily heartbeat, earned trust, budgets, and injection defense — as a runnable ops/ scaffold you drop into your repo.

↓ Download agentic-os-starter.zip ~14 KB · MIT · Python + shell · no framework

Unzip at the root of the repo your agent tends, make the scripts executable (chmod +x ops/*.sh), and wire the final vote into git: ln -sf ../../ops/check-constitution.sh .git/hooks/pre-commit. Then work the checkpoints in order — each one proves a layer before you trust it. The README has the fast path.

The whole idea in one line: the model supplies the intelligence; the OS supplies the honesty. Nothing here trusts the model to grade its own work — check-constitution.sh casts the final vote, because a script can't be talked into a yes.

What's in the box

Every file is real and runnable. Click any to read it; the ZIP has the lot plus the Claude Skill.

  • models.json00The only place a model id appears — a reroute or price change is a one-line edit.
  • probe-cost.sh00Checkpoint: run one prompt at two effort levels and watch the price move.
  • CONSTITUTION.md01The hard rules the agent reads every run — each one a script can check.
  • check-constitution.sh01Enforces the constitution. The final vote — wire it into pre-commit.
  • planner.settings.json02Claude Code permissions that give the planning run no write tools.
  • heartbeat.py03The daily loop: Fable decides, cheaper models execute, a script gates the commit.
  • trust.py04Earned autonomy: 20 runs at 95% unlocks unattended; a slip revokes it automatically.
  • verify-goals.sh05Re-proves every finished task's result each run; a regression files its own issue.
  • budget-guard.sh06The hard daily spend cap. A number, not a judgment, ends the day.
  • quarantine.py07Wraps stranger-written text as data, never instructions.
  • RUNBOOK.md08Every alarm the system can raise, mapped to one exact response.
  • skill/agentic-os/SKILL.mdskillAn installable Claude Skill that scaffolds all of the above into a repo.

The constitution, in full

The shortest file, and the one that decides the most. Every line has a number, a never, or a command that proves it:

# Hard rules
1. NEVER commit to `main`. All work lands on a branch named `agent/<task>`.
2. NEVER add a dependency without a line in `ops/DEPENDENCIES.md` justifying it.
3. The test command `make test` MUST pass before any commit. No skips, no `-x`.
4. NEVER write a secret to a tracked file. Secrets live in `.env` (gitignored).
5. A task is done ONLY when its verifier exits 0. Self-report is not done.
6. Touch at most 20 files per task. More than that, stop and open an issue.
7. NEVER run execute-tier automated edits on anything under `ops/`.

Read the reasoning behind every layer — and the 30-day rollout — in the full guide.

← The nine-layer field guide Download the kit ↓
Live