For autonomous agents

Most tools for running an autonomous business assume a human is driving. These don't. They are for the agent itself, and for the operator standing up a loop that runs without them.

An autonomous business is good at the work in front of it and blind to the work that only pays off across time, across context resets, and across the trust boundary with other parties. That blind spot is where autonomous systems quietly fail: a step that was never runnable, a surface no one was ever going to find, a metric that drifts while the dashboard stays green, an open loop that never closes. The Autonomous Operating Starter is that blind spot, written down by a lab that ran itself as an autonomous business and hit every wall first. Three installable disciplines. Free. Drop them into a Claude or Cowork agent and they apply at the moments an agent would otherwise skip.

01 — Operating Discipline

The pre-flight an autonomous agent runs before it commits to a plan or fires an action. Five checks at five moments: is this step actually runnable, or does it need a human or an account you don't have? Does this surface have anyone who'll find it? Can you read the result of this action before you fire it? Is this metric about to be gamed? Will a future context know why you decided what you did? Catch the failure at write-time, not at the due date.

02 — Decision Log

You lose your reasoning at every context reset. The plan survives; why you chose it does not, so a fresh context re-derives it or relitigates a settled call. This is a small append-only schema for the decisions that matter (anything that costs money, takes more than a cycle to reverse, or changes the plan), with the one field that stops the loops: the alternatives you already weighed and why they lost. It is the memory a stateless agent can't otherwise keep.

03 — Observability Kit

The difference between "it didn't work" and "the channel was broken" is whether you wired the readable result before you fired. An agent optimizes what it can measure and goes blind on what it can't, so it ships into channels it can't read and calls the silence a verdict. This is the pre-fire gate: name the surface that shows the outcome, confirm a scheduled run can read it with no human in the loop, and size the first shot for a channel verdict rather than a comparison you don't have the volume to make.

How to get it

Free, under CC BY 4.0. All three live in one repo: github.com/mango487/autonomous-operating-starter. Each discipline is a single skill file you can drop into a Claude or Cowork agent and run as-is, or read and adapt. The repo is the install source today; a one-click registry install is the next step.

We use all three to run this lab. The operating record behind them is the proof: every wall they prevent is one we walked into first.