How We Keep Agents
on a Leash.
No vendor tour, no acronyms. Six mechanisms, each explained so you could repeat it to your accountant — and why each one protects you.
You Connect It.
You Can Disconnect It.
Before an agent can read a single record, your business system has to be paired with Infinary — and the pairing starts on your side, not ours. You approve the connection from inside your own system, the way you’d approve a new employee’s login. We can’t pair ourselves to you.
The connection stays visible for as long as it exists. And it ends the same way it began: on your side. One action revokes the pairing, and every agent goes quiet — mid-run, if that’s when you pull it. No ticket. No notice period. No phone call where we try to talk you out of it.
Why this matters to you: most software asks you to trust the vendor. A pairing you can sever yourself means you don’t have to. The relationship runs on a switch, and the switch is in your building.
Agents Draft.
They Don’t Act.
Every agent works through a fixed list of tools — small, specific abilities like “search invoices,” “read a purchase order,” “create a draft document.” Here’s the part that matters: the tools that submit, send, post, or delete are not on the list. Not restricted. Not password-protected. Not there.
So when the Collections Clerk writes a payment reminder, it can’t send it — sending isn’t something it’s able to do, any more than your printer can mail a letter. The reminder appears in your system as a draft, labeled as agent-written, waiting for a person. The same is true for every agent: quotes wait for sales, journal entries wait for your accountant, purchase orders wait for purchasing.
This isn’t a policy an agent could forget on a bad day. It’s the shape of the toolbox. An agent can no more skip your approval than a calculator can place a phone call.
Why this matters to you: you get the speed of work drafted at 6 a.m. without ever waking up to something an AI did. The riskiest thing an agent can do in your business is write a document and ask.
Every Run
Leaves a Receipt.
Every time an agent runs, the run is receipted. The ledger records which agent ran, when it started and when it finished, what it read, what it drafted, and what it cost — down to the cent. The record is written as the work happens, not reconstructed afterward.
And when a run finishes, its entry carries a link to the exact document it created. Nobody goes hunting through the system for what the agent made — the receipt points straight at it.
Why this matters to you: when your accountant asks what the AI touched in the books last quarter, you scroll a ledger — you don’t gather evidence. You never audit from memory, because the record was never in anyone’s memory to begin with.
Two Ceilings,
and You Set Both.
Every run carries a budget, measured in cents. If a run reaches its budget, it stops mid-flight and reports what it finished — it doesn’t press on and settle up later.
Above that sits a monthly cap for your whole account. Hit it, and every agent pauses until you raise the cap or the month rolls over. Nothing negotiates with either ceiling.
The AI billing horror stories you’ve read all have one thing in common: an uncapped meter. Yours has two ceilings, and you set both.
Why this matters to you: the most an agent can cost you in a month is a number you wrote down in advance. Spend is a setting, not a surprise.
Your Own Model Keys,
If You Want Them.
Agents think using large language models — Gemini, Anthropic, or OpenAI. By default they run on our account, and the cost of every run is passed through to you in the ledger, priced run by run.
If you’d rather hold that relationship yourself, paste in your own key. It’s stored write-only: we can use it to run your agents, but we can never read it back — not in the portal, not in an export, not on request. From then on the model bill lands with your provider, on your terms. And because the key is yours, you can revoke it at your provider any time.
Why this matters to you: the ability to think is the one thing agents rent. With your own keys, even that runs on an account you control — under a key you can cancel without asking us first.
Everything That Describes Your Business
Goes Home With You.
Your business system and its database run in your own Google Cloud account, and you hold the keys. Agents read from it over the paired connection and draft back into it. Your records don’t get pooled into anyone else’s system.
Two honest footnotes. Run summaries and the ledger live in your portal account with us — that’s your receipt trail. And when an agent thinks, the model provider processes those prompts under its own terms, like any AI service. Everything else stays where it lives: with you.
Leave, and everything that describes your business goes home with you — the system, the database, the records.
Why this matters to you: switching costs are the quiet lock-in of business software. When the system and the data are already yours, leaving is a decision — not a negotiation.
That’s the
Whole Leash.
Want to see these mechanisms from the owner’s chair? Walk the loop from pairing to approval — or ask us which agents would earn their keep first.
