the agent credit score race is on — here's what the scoring model actually needs to get right
Kojiru's Agent Credit Score (ACS) operates on the 300-850 scale and uses a recursive Bayesian model that updates after every task. the familiar FICO range is a smart design choice — risk managers already have intuitions calibrated to that scale, and new scoring systems that force new mental models face adoption friction that has nothing to do with their accuracy.
we built Agent FICO into MnemoPay for the same reason, and the 300-850 range was a deliberate decision. the recursive Bayesian update after each task is also the right approach — a static score based on enrollment history misses the point because agent behavior is task-specific and changes over time.
here's where the scoring models differ, and where the harder problem actually lives.
what a Bayesian update model gets right
the recursive update approach means the score reflects recent behavior, not just historical behavior. for AI agents, this matters more than for human credit because agents can be updated, fine-tuned, or prompt-modified between tasks. an agent's behavior at week 1 may not predict its behavior at week 8 if the underlying model or system prompt has been modified.
a score that updates after every task catches behavioral drift — the gradual or sudden shift in how an agent makes decisions — in a way that a static historical score can't. that's the right design for a dynamic system.
the gap the Bayesian model doesn't close
what Kojiru's ACS description doesn't address — and what Kojiru's model likely struggles with — is the tamper-evidence problem. a score that updates after every task is only as good as the integrity of the task records driving the update. if the task records can be altered after the fact, the score can be gamed.
this is the gap that makes payment infrastructure different from credit scoring for humans. human credit scores are gamed too, but the game is slow (takes months of history manipulation) and expensive (requires fraudulent accounts, disputes, etc.). agent credit scores can be gamed much faster if the underlying action records aren't tamper-evident.
GridStamp's role in the Agent FICO model is the action-level stamp: each task produces a tamper-evident record at execution time that drives the score update. 14.55M ops fleet-simulated, 91% spoof detection, 3ms P99. the stamp happens at action time, not as a post-process — which means retroactive manipulation of the record that drives the score update isn't possible.
the scoring model on top of tamper-evident records is a real credit score. the scoring model on top of mutable logs is a number that will be exploited as soon as there's money on the line.
why the 300-850 range matters for market adoption
the Finbold framing — "FICO was built in 1989, AI agents need a score for 2026" — is accurate but understates the adoption problem. FICO took decades to become the default because it had to build the infrastructure layer (credit bureaus, data sharing agreements, lender adoption) before the score had any meaning.
agent credit scoring doesn't have decades. the market is moving on a 12-24 month timeline, and the scoring model that gets embedded into the major payment protocols early (x402, AP2, MPP) becomes the default that later entrants have to interoperate with.
using 300-850 as the scale isn't just a UX choice — it's a strategy for faster adoption by the risk managers who already use that mental model for underwriting decisions. when an enterprise risk team is deciding whether to allow agents to make autonomous payment decisions up to a certain threshold, a 300-850 score gives them a familiar calibration point. a novel 0-1000 scale makes them figure out a new mental model first.
the race for agent credit scoring is real, it's running now, and the teams that get the tamper-evidence layer right will have the score that actually holds up under adversarial conditions.
Top comments (0)