FICO was built for humans who pay once a month. AI agents pay once a millisecond.
Wayne Faulkner's framing is right: FICO's genius was giving lenders a single number they could trust without reading every bank statement. The model held for 37 years because humans operate on monthly billing cycles, have one SSN, and accumulate history slowly. AI agents do none of those things. An agent handling autonomous e-commerce can execute 300 transactions before a human even opens their laptop. FICO can't score something that settles faster than it can report.
The question isn't whether agents need a credit score. They obviously do. The question is what the inputs look like when you're scoring a non-human entity that has no birth certificate, no bank account in its own name, and a behavioral signature that rewrites itself every time the underlying model gets fine-tuned.
the cold-start problem is worse for agents than it was for immigrants or 18-year-olds
FICO's original cold-start problem — how do you score someone with no credit history? — was solved by secured cards, credit-builder loans, and becoming an authorized user on a parent's account. The workarounds took months and required a human with a social graph willing to vouch.
An agent has none of that. It spawns with zero history, gets handed a spend policy by a developer, and immediately starts operating in production. The first transaction it makes is also its entire credit history. If it misbehaves on transaction one, there's no prior good behavior to offset it.
There's a second problem: identity drift. A human's identity is anchored by SSN, biometrics, and a legal name. An agent's identity is anchored by... what exactly? A wallet address that can be regenerated? A model fingerprint that changes after every fine-tune? A session token that rotates hourly? The scoring system has to work even when the underlying entity is non-deterministic.
what machine-native scoring actually needs to track
After running 14.55 million ops through GridStamp's fleet simulation, a few inputs emerged as genuinely predictive — none of which FICO uses:
Settlement velocity. How fast does this agent close what it opens? An agent that consistently settles obligations within 50ms of the agreed window is a different risk profile than one that leaves pending transactions open for hours. Humans can't produce this signal — agents generate it constantly.
Blast-radius consistency. Does the agent stay within the scope of its declared intent? An agent authorized to purchase API credits that also starts spinning up compute instances has demonstrated scope creep — a behavioral anomaly that precedes most agent-related fraud events we've seen in the fleet data.
Reversibility ratio. What fraction of the agent's actions are reversible vs. permanent? An agent with a high permanent-action rate operating at high frequency is a governance time bomb. Scoring this catches problems before they compound.
Cross-session identity continuity. Does the behavioral fingerprint stay stable across sessions? Sudden style shifts in tool-call sequencing, unusual timing distributions, or new permission requests without corresponding task context are all signals that something changed in the underlying model — and the downstream trust score should reflect that.
These are the inputs GridStamp's Agent FICO tracks — same 300-850 scale Faulkner referenced, updated with every settled charge rather than every monthly cycle.
the governance layer that makes scoring actionable
A score is only useful if something enforces it. FICO works because lenders actually check it and have systems that gate on it. Agent credit scoring needs the same enforcement layer built in from the start — not bolted on after an incident.
GridStamp ships three enforcement primitives on top of the score:
First, per-agent spend limits that update dynamically with the score. An agent that just scored its first 40 transactions cleanly gets a small limit increase without human review. An agent that shows blast-radius anomaly gets throttled automatically.
Second, cryptographic audit logs on every action — not a database write, a Merkle-anchored ledger entry. When a regulator or auditor asks "what did this agent do and who authorized it," the answer is a tamper-evident chain, not a log file someone could edit.
Third, human-in-the-loop escalation for actions that cross configurable trust thresholds. The system doesn't block the agent — it pauses and routes to a human approver. 91% spoof detection rate in stress testing means the escalations that fire are real anomalies, not noise.
GridStamp hit 221 tests green before shipping the first public beta. The SDK is free. The hosted Agent FICO API is $49/month — roughly what one hour of undetected agent misbehavior costs in token burns.
where this goes
Faulkner's piece ends with the observation that agents need "machine-native credit scoring." The harder follow-on question is: who issues the score?
FICO's monopoly came from network effects. Lenders trust it because every other lender trusts it, which means the data pool is the largest available. Agent credit scoring needs the same convergence to a shared standard — not ten competing scores that no one can compare across providers.
The infrastructure to get there exists now. What's missing is adoption by the platforms that are already wiring x402 payments and MCP server access into production agents. Those platforms — Coinbase, Stripe, AWS AgentCore — are the lenders of the agentic economy. They're the ones who need to check the score.
If you're building agent infrastructure that touches payments, authorization, or multi-step execution, the governance primitive has to come before the go-live, not after the first incident.
Top comments (0)