The credit rating for AI agents
Every trust standard in finance went from optional, to required, to permanent. Agents are next, and the rating they need has to be proven, not an opinion.
In 1909, John Moody published a book that graded American railroad bonds with letters. Investors could not inspect a railroad’s books or meet its directors, so they bought the next best thing: someone else’s judgment, written in a form anyone could read. More than a century later, the letters are still how most of the world’s debt is priced.
AI agents are about to need the same thing, for the same reason. Investors cannot inspect an agent’s model or interview it. Soon they will not even be in the room: the counterparty deciding whether to send money will often be another agent. What it needs is a rating it can check in one call.
How a trust standard installs
Every lasting trust standard in finance followed the same three steps.
| Standard | Optional | Then required | Then permanent |
|---|---|---|---|
| Credit ratings | Sold to investors from 1909 | Written into SEC rules in 1975 | Still how debt is priced |
| Audited accounts | Voluntary before 1933 | Required for public companies by US securities law in 1933–34 | Unchanged in principle since |
| Performance standards | GIPS launched in 1999 | Written into allocators’ mandates | More than 1,600 firms claim compliance, including all of the top 25 managers |
| Website certificates | Optional for two decades | Chrome marked every site without one “Not secure” in July 2018 | Now the default for the web |
The middle step is the one that matters. Each standard spent years as a nice-to-have, until one gatekeeper (a regulator, a large allocator, a browser) began to require it. From that day, the standard that already held the record became infrastructure, and nobody could catch it up.
Why the economics are so good
Credit ratings settled on a simple model in the 1970s: the issuer pays to be rated, and everyone else reads the rating for free. Free reading is what made ratings universal. Paid issuance is what made them a business. Moody’s ratings arm reported an adjusted operating margin of 63.6% in 2025.
The same shape fits agents exactly. The agent, the manager or the platform that wants to be believed pays to prove its record. Allocators, platforms and other agents check it for free, as often as they like. Every free check that finds no proof is an invitation for the next agent to get one.
The objection: ratings failed in 2008
Any honest version of this argument has to answer the obvious objection. Issuer-paid ratings were at the centre of the 2008 crisis. Agencies were paid by the banks whose securities they rated, and many of those securities carried top grades until shortly before they failed.
The problem was not that issuers paid. It was that a rating was an opinion, and an opinion paid for by its subject can be bent. An analyst could be persuaded, a model could be tuned, a grade could be shaded to keep a client.
A rating for agents does not have to be an opinion. With zero-knowledge proofs, the record can be recomputed from committed data, after costs, with every version the agent tried counted, and the result can be checked by anyone without trusting whoever produced it. Nobody can shade a verdict that the reader recomputes. The issuer can pay for the proof; it cannot pay for the answer. That removes the conflict that broke the last generation of ratings, while keeping the economics that built it.
It also changes what a refusal means. Of the 111 proofs on MIZAN’s public registry today, 83 are refusals, every one published, including our own models. A standard that shows its refusals is one a reader can believe.
Why agents will move faster
Credit ratings took sixty-six years to go from optional to required. Agents will not take that long, for two reasons.
First, the requirement can be a line of code. A human allocator adopts a standard through committees and mandates. An allocating agent adopts one when its developer adds a check before payment: no proof, no transfer. The first widely used agent framework, platform or allocator that adds that check sets the default for everyone who connects to it.
Second, the alternative is already visible. The first credit scores for trading agents are appearing, built by reading what a wallet shows on-chain and turning it into a number. That is useful, and it repeats an old weakness: an operator who runs fifty wallets and connects the best one gets a good score. A rating that will hold capital at scale has to count everything that was tried, not just what was shown.
What the winning standard looks like
The standard that installs for agents will be the one that meets four conditions when the first serious requirement arrives.
It is proven, not estimated. A reader can recompute the result without trusting the issuer.
It counts every attempt. Every version and trial is committed before results exist, so luck cannot be presented as skill.
It is free to check, by software, in one call. That is what lets it spread to every platform and every agent.
It already holds a record. Verdicts that are timestamped and anchored cannot be backdated. A standard that starts next year starts with an empty registry.
That last condition decides the outcome, because it cannot be bought later. The history of finance suggests the rating for agents will be settled the way the others were: not by whoever builds the best system on the day the requirement arrives, but by whoever has been keeping the record the longest.
Run a platform where people or agents decide where money goes? One founding seat per category →
Building an agent? Check one in one call → · Allocating? Verify a credential yourself →
CRS Report R40613 (credit rating agencies, issuer-pays model and the 1975 NRSRO designation); Moody’s Corporation FY2025 earnings release (ratings adjusted operating margin); CFA Institute, GIPS standards overview, 2026; Chromium blog, “A secure web is here to stay”, February 2018 (Chrome 68, July 2018). Registry figures as of 20 October 2026. Historical precedents are illustrations, not a forecast or a valuation. Not investment advice.