FICO was built in 1989. AI agents need a score for 2026. here's what that score actually measures

when agents need to pay for API calls or settle vendor invoices, they hit a wall. no bank account. no FICO score. no credit history.

that's FinBold's framing, and it's accurate as far as it goes. but the more interesting question isn't "why don't agents have a score" — it's "what would you actually measure if you were building one from scratch in 2026?"

FICO built their model around the data available in 1989: payment history from credit bureau files, credit utilization, account age. those dimensions were proxies for a single underlying question — is this person likely to repay what they borrow?

the agent version of that question is different: is this agent likely to complete its authorized task, within its authorized scope, without going sideways?

the five dimensions of agent behavioral credit

an Agent FICO built for the x402/MCP ecosystem tracks five dimensions:

1. settlement velocity — how fast does the agent settle charges after task completion? agents that initiate payment but delay or fail settlement are high-risk
2. task completion rate — ratio of task-initiated sessions to successfully-completed sessions; agents that start work and abandon are a liability
3. capital-to-output conversion — how much did the agent spend, relative to the value of what it produced? a high-spend, low-output pattern is anomalous
4. scope drift — did the agent operate within its authorized permission scope? exceeded scope on even one transaction is a red flag
5. dispute frequency — how often does the agent's counterparty dispute the charge or the output?

those five dimensions produce a 300–850 score. 300 means no history or active flags. 850 means clean history across hundreds of settled transactions. the score updates with every settled charge.

how it works in practice

MnemoPay ships Agent FICO as part of the payment SDK. you don't need a separate API call to score an agent — the score travels with the payment event. API operators set a minimum threshold (e.g., minFico: 650) in their gateway config. agents below threshold get a 402 response indicating insufficient trust history, not insufficient funds.

new agents start at 300 and ramp quickly — 30 settled transactions at $0.001 is enough to push a clean agent above 600. bad actors' scores decay on dispute. the feedback loop is tight enough that it self-regulates without manual intervention.

672 tests cover the scoring model, EWMA anomaly detection, and the merkle ledger that makes every score audit-traceable. that last part matters for Article 12 compliance.

the infrastructure gap FinBold didn't mention

FinBold's piece correctly identifies that agents need a reputation system tracking "uptime, settlement velocity, capital-to-output conversion." what it doesn't mention: that data also satisfies EU AI Act Article 12 record-keeping requirements. tamper-evident behavioral records over the agent's operational lifetime — that's what Article 12 requires, and it's the same data the scoring model runs on.

building Agent FICO compliance infrastructure doubles as audit trail infrastructure. the ledger is the same.

the SDK is free at mnemopay.com. hosted Agent FICO API is $49/mo.