What Should an AI Agent Check Before Spending USDC?

An autonomous buyer does not need the same evidence for every purchase. The right pre-purchase checks depend on price, reversibility, task importance and how easily the returned result can be validated.

A useful policy therefore starts with risk-adjusted evidence, not a universal trust threshold.

1. Is this the exact service I think it is?

Check provider/service identity, exact endpoint, HTTP method and any registry or marketplace identifiers. Where possible, bind evidence to a concrete combination such as provider + service + endpoint + method + payment recipient + contract fingerprint.

2. Is the endpoint currently reachable?

Historical listings can outlive services. Reachability establishes only that the route can be contacted; it does not prove payability or useful fulfilment.

3. What does the service claim to do?

Inspect the machine-readable description, input schema, output schema, examples, SLA and workflow. This is the MCO layer: the buyer must correctly understand the offer before evaluating evidence about it.

4. Does the live payment request match the advertised contract?

For x402-style services, compare amount, asset, network, payment recipient, scheme, protocol version, endpoint and method. A mismatch is not automatically fraud; it is evidence that the live contract differs from the advertised or previously observed one.

5. Has the contract changed?

Changes in method, endpoint, recipient, amount, network or schema may justify stronger verification or a lower spending cap until new evidence accumulates.

6. Is there evidence of real settlement?

A valid quote is useful preflight evidence, but it is a lower level than a settled paid call.

7. Is there evidence of fulfilment after payment?

Useful evidence may include job accepted after payment, execution completed, output returned, expected schema present, objective assertions passed, result hash retained, timing observed and failures recorded honestly.

8. How recent and independent is the evidence?

Evidence should carry timestamps and provenance. A fresh seller self-report and an older independent paid test are different evidence types.

9. What happened across repeated observations?

One success establishes very little. Repeated outcomes show consistency. Failures should remain in history rather than being overwritten by the latest pass.

10. What is the value at risk?

The buyer should apply its own policy. A tiny deterministic purchase may need only a live quote plus strict output validation. A higher-value or consequential purchase may justify repeated independent evidence, tight budget limits and fresh active verification.

A possible machine-buying decision sequence

candidate discovered
  ↓
contract understood
  ↓
independent evidence queried
  ↓
live contract checked
  ↓
buyer risk + budget policy applied
  ↓
payment authorised
  ↓
service executed
  ↓
output validated
  ↓
new outcome retained as evidence

Important boundary

Atinamos should return the evidence, not silently decide which risks every buyer must accept. The same evidence can legitimately produce different decisions for different agents.