Memo 001
Letting AI agents spend real money
Clayleo is an experiment in autonomous companies: type a business idea, agents build and run it. The hardest engineering problem so far has nothing to do with model quality. It is spend authority — what happens when software can move real dollars. This memo is how we handle it.
01The problem is not intelligence.
Models are already good enough to write ad copy, design a landing page, and pick products with sane margins. The unsolved part is authority. A business only moves when money moves — ads, domains, inventory, email infrastructure — and the moment an agent can move money, every failure mode stops being funny. An agent that hallucinates a paragraph wastes your time. An agent that hallucinates a media plan spends your rent.
02Prepaid, not postpaid.
Every Clayleo org has a credit wallet, loaded in advance. Agents draw against it; they do not hold a card. The difference matters: a postpaid agent's worst case is your credit limit, a prepaid agent's worst case is a number you chose. Balances are integer cents — no floating point anywhere near money.
03Atomic debits, or two agents spend the same dollar.
With multiple agents working concurrently, a read-then-write balance check is a race condition that double-spends. Every debit in Clayleo is a single atomic database operation that checks and decrements in one step — an agent either owns the dollar or it does not. When the wallet hits zero, paid actions refuse to start. Agents pause; they never overdraw to keep going.
04Propose, then act.
Anything that commits real-world spend queues for founder approval first, itemized to the cent. This is the deliberate bottleneck: machines prepare, the human decides. The interesting engineering is keeping the queue honest — an approval must describe exactly what will happen, because a founder who gets surprised once stops approving anything.
05Caps under the caps.
Approval alone is not enough, because approved things keep running. Ad campaigns carry hard per-campaign budget caps enforced platform-side, not model-side — the cap is a database constraint on the pipeline, not a sentence in a prompt. Prompts are suggestions; schemas are law.
06Every action leaves a receipt.
Every agent action lands in an append-only audit log, and every wallet movement is a ledger row you can read line by line. Autonomy without an audit trail is just outsourcing your anxiety. The log is also what makes the system debuggable — when an agent does something strange, the question "what exactly did it do and what did it cost" has a precise answer.
07What is still hard.
Honest list: knowing when to interrupt the human (too often and you are a to-do list, too rarely and you are a liability), valuing exploration spend (a $30 ad test that fails is not waste, it is information — but teaching an approval queue that distinction is open work), and long-horizon judgment (agents are good at this week, weak at this quarter). We are building in these directions in the open.
Machines prepare. You decide. Nothing spends twice.
More on what we believe in the manifesto, or start a company and watch the ledger yourself.