Before You Let an AI Agent Move Money

AI systems are moving from recommending actions to participating in commerce. Google describes UCP as an open standard for agentic commerce and AP2 as a compatible authorization layer built around typed mandates, guardrails and an auditable proof of intent. OpenAI and Stripe have separately developed the Agentic Commerce Protocol for transactions between users, agents and businesses.

This infrastructure matters because a new boundary is appearing: an agent may not merely suggest a purchase. It may help initiate, negotiate, authorize or execute a consequential economic action.

Anthropic's Project Deal offers another signal. In an experimental marketplace, AI representatives negotiated real items on behalf of people and closed 186 deals with a total value above $4,000. The experiment was not a production payment network, but it demonstrates that delegated economic action is no longer a purely theoretical interaction pattern.

Consider one ordinary distributed-systems failure.

An agent is instructed to pay a supplier. The payment request is sent. The network times out before the agent receives a definitive response.

If the payment failed, retrying may be correct. If it committed and only the acknowledgement was lost, an uncontrolled retry may create a duplicate payment.

The second request can return 200 OK. The agent can report success. The final state can even look superficially plausible. Yet the path may already have violated the business invariant.

Authorization solves only one part of trust.

Protocols such as AP2 are valuable because they make delegated intent and spending constraints more explicit. Google's developer guidance describes an intent mandate, a payment mandate bound to a specific cart and amount, and a receipt that closes the authorization trail.

But a production application still has to preserve correctness across its own state machine, provider behavior, retries, concurrent changes and recovery logic.

For a consequential financial workflow, we want answers to eight questions:

The trust contract should survive the model.

A useful financial-agent contract is not “GPT/Claude/Gemini should behave carefully.” It defines conditions that remain testable regardless of which planner selected the action.

INTENT
User approved supplier payment ≤ $1,000

STATE
Invoice is unpaid

TIME
Authority remains valid at execution

TRANSITION
UNPAID → PAYMENT_PENDING → PAID

RECOVERY
Unknown commit triggers reconciliation, not blind retry

VERIFICATION
Provider state and authoritative ledger agree

EVIDENCE
Intent + authority + attempts + reconciliation + decision survive

Now the engineering question becomes sharper:

A safer ambiguous-failure path.

That path is not universal. Some payment systems provide idempotency primitives; some expose reconciliation differently; some business processes have delayed settlement or multi-party authority. The point is not that RESONANCE has invented a universal payments algorithm.

The point is that the recovery path itself needs to be explicit, testable and evidenced.

And this is where we need the market.

We can generate synthetic failure cases forever. Real production constraints will teach us more.

If you submit a useful case, our first goal is not to sell you a platform. It is to isolate the missing guarantee, map the trajectory and return one concrete verification hypothesis. If multiple independent teams reveal the same missing capability, that becomes evidence for what should be built next.

What happens to the answers?

Meaningful submissions become structured Problem Cards: workflow → failure → impact → workaround → missing capability → acceptance condition. They are scored as product signals and clustered in a Demand Graph. Synthetic examples are marked as synthetic and excluded from real market metrics.

Article #005 will not be called “market-selected” until the evidence threshold is reached. The current rule requires at least ten meaningful external workflows and at least three real cases in the winning cluster. Until then, the next topic remains open.

Don't ask the market whether it likes our product. Build a conversation in which the market reveals the product it needs.