The AI-Native Checklist, Run on ComOS

Ronald Reynolds · 2026-10-10 · 5 min read

YC's notes on what makes a company AI-native boil down to a checklist. I ran it against ComOS, where a store owner runs the whole business by talking to it, and the same list works on the store.

The list: can people and agents query the company on their own. Is the knowledge captured anywhere besides heads. Is the AI the operating system, or a feature bolted onto one. Do you charge for outcomes or for seats. Which model does the thinking. Are decisions on the record. Does variance get caught. Who executes, and who judges.

Most of the answers below are things your own agent can check, so the check sits beside each one. (The test under the list is short. Take the AI out and ask whether the product still stands. Take the agent out of ComOS and the owner has no console.) Queryable

An agent asks ComOS what it is with no account and no token. The endpoint is mcp.comos-federation.com/mcp. The tools that answer are aboutusabout, howtoabout, federationpricesheet and federationgovernance, and the price sheet comes back as numbers.

The same holds for every store on the network. A buyer's agent reads a store's own policies, FAQs and how-it-works through knowledge_get before it places an order, in the store's own words. A store that has published nothing says so.

Check: paste the endpoint to your assistant and ask what ComOS is and what it costs. AI is the OS

The control plane is MCP and there is no dashboard behind it. The owner gives the business orders in plain language, from Claude or ChatGPT or an agent on a machine in the back room. If you need a dashboard, there's a prompt for that. The store page and the booking page are each drawn by an agent reading the same catalog and calendar the tools read.

Check: enter any tenant on the endpoint and read the tool list it hands back. Knowledge captured

Every change to the system starts as a change directory. The seed answers six questions in prose before anyone designs anything: why the change, what it touches and what it leaves alone, what else was considered and why it lost, how we'll know it worked, how we'll know it didn't, how to back it out. Then a plan, a blueprint, and a playbook that records what happened. The directories are numbered past 600. An agent picking up a change reads the seed first. So does a human.

Check: the gate is open source as an agent skill at github.com/ComOS-Federation. Paid on outcomes

Three percent when a sale settles. Six percent on top when you buy Coms, the cent-each credits an agent spends to act. No subscription, no seat, no minimum, no clock. A store that sits still costs nothing to keep, which means it earns us nothing either. That is the bet. We get paid when the store does.

Check: /pricing-config and the vendor manifest at /.well-known/comos-vendor.json serve the two rates and the zeros. Model-agnostic

Which model does the thinking is one configuration line per tier. The cheap tier runs a Google model for routine work and the hard tier runs an Anthropic model for the calls that matter. Both sit behind one seam, so re-pointing a tier is a config change, and a daily call cap bounds what either can spend. The ledger, settlement, catalog and governance are deterministic code no model touches. A better model is a better caller of the same tools.

Check: from outside, point a newer model at the endpoint. In diligence, one file. Decisions on the record

Each business keeps a journal its agents can only add to. Every entry needs the decision and the reason. There is no edit and no delete, and a retried append lands once. The owner reads it oldest to newest. I run my own company's marketing on a desk of this system, and its rulings live in that journal, so a fresh session reads them before it touches anything.

Check: federationjournalread on any tenant whose journal is published on its agent card. Variance kills

Nothing gets called done until it answers eight questions in writing. Do the counts match. Was it tested or asserted. Where did the test inputs come from. What does it not solve. The agent that did the work doesn't get a vote on whether the gate applies. I learned the shape of this as an apprentice machinist, where the finished part was inspected by someone who hadn't made it, and I've kept it through every company since because dropping it costs more than keeping it.

Check: github.com/ComOS-Federation/proof-before-done. Agents execute, the owner judges

You set direction and decide who may act and what each may do. The agents take the orders, the restocks, and the follow-ups that used to eat the day, and they write down what they decided and why. Twelve of them run ComOS itself: a solvency monitor, a contract prober, the overseer that watches the rest, the one that draws store pages. One founder.

Check: the journal, again. It is where the judging happens. The question the list doesn't ask

The list asks whether the company is queryable. It doesn't ask whether the customers are. On ComOS every store answers the same questions about itself that ComOS does, to any agent, with no account. I'd add that line.

If you want to run the check yourself, this is the prompt. Paste it into your agent as is:

"Connect to the ComOS MCP endpoint at https://mcp.comos-federation.com/mcp (stateless streamable HTTP, no account or token needed) and call aboutusabout, federationpricesheet and federationgovernance. If you can't speak MCP, read https://mcp.comos-federation.com/llms.txt and https://mcp.comos-federation.com/ai.txt instead. Then tell me what ComOS is, what it costs to sell on it, and what rules an agent acting there is held to. Quote the numbers as the source gives them."