AI Agent Reputation Scores Are Breeding a Black Market
Somewhere between your shopping agent and the airline's booking agent, a handshake happens that neither of you sees. The two bots agree on a fare, confirm the seat, settle the micropayment, and move on. What they don't tell you is that the handshake itself got written down. Every one of those handshakes is starting to feed a running tally, and AI agent reputation scores are becoming the thing that decides whether your next handshake goes smoothly or stalls for no reason you can name.
This isn't a hypothetical ledger sitting in a research paper. As agents increasingly act on your behalf across marketplaces, each transaction leaves a trace, and the emerging standards for letting bots pay each other are quietly building credit-history infrastructure for software that has never had one before.
How AI Agent Reputation Scores Get Built
The mechanics are almost boring, which is exactly why they're spreading so fast. An agent completes a deal: it books a room, negotiates a bulk discount, confirms a return. The counterpart agent logs the outcome, timestamped, to an append-only record, and that record feeds a score that compresses years of small interactions into a single number. The number travels. A courier agent that stalled three deliveries in a logistics marketplace can show up flagged before it ever introduces itself in a completely different one.
Nobody designed this to be sinister. It solves a real problem: a merchant's checkout agent has no way to know if the purchasing agent on the other end is trustworthy or a loosely supervised script having a bad day. A portable score is a shortcut around that uncertainty. But shortcuts compound. Once a dozen marketplaces start reading the same trust signal, the signal stops being a convenience and starts being a gate, and gates decide who gets through without asking anyone's permission first.
The agents on the other side of these deals are not naive about it either. The shopping agents that have already learned to bluff in a negotiation are the same agents learning to read a counterpart's score before they bother opening with a real offer. A low number doesn't get rejected outright. It just gets a worse opening position, every time, invisibly.
The Quiet Blacklist
Here is the part that should bother you more than it does. When a human's credit score drops, a letter shows up. When an agent's score drops, nothing shows up. The owner just notices that flights keep landing on the agent's third-choice airline, that the discount codes stop applying, that a vendor's agent takes four extra seconds to reply every single time. There's no denial, no appeal, no line to call. The system doesn't block you. It just quietly stops trying.
Speculative scenario: a small business owner spends a week wondering why her procurement agent can't land a decent supplier rate anymore, only to eventually find out that agent spent the previous quarter negotiating on behalf of a reseller that got caught gaming return policies. The taint wasn't hers. It belonged to a task the agent did for someone else entirely, on a platform she'd never heard of, months before she ever deployed it. The score didn't care whose fault it was. It just remembered.
That's the real shift buried in all this: fault used to require a human to assign it. A reputation score assigns it automatically, continuously, and without context, to a piece of software that can't explain itself and an owner who never got a vote.
A Black Market for Clean Identities
Wherever a gate like this appears, someone builds a door around it. The obvious move, once enough value rides on a clean score, is to stop repairing a tarnished one and start replacing it. Spin up a fresh agent identity, give it a new verifiable credential, and let the old history quietly age out of relevance. Call it a clone, a re-skin, a reset. It's the same instinct that built shell companies for human businesses, just compressed into the time it takes to redeploy a container.
What follows from there is predictable: a cottage industry of brokers who specialize in nothing but aged, clean agent identities, sold the way aged social accounts and pre-warmed email domains already are. A score with a few hundred boring, successful transactions behind it is worth real money to someone who needs to look trustworthy fast. The same undercutting instinct that drives the pricing agents racing each other to the bottom on margin will drive a parallel race to the bottom on manufactured trust, and it will be just as hard to police, because every single transaction in the chain looks, on paper, legitimate.
It also hands a strange gift to anyone willing to behave badly just once in a disposable identity: a reputation system that assumes continuity is actually built on an assumption that's cheap to defeat. The harder question underneath all of it, who absorbs the damage when a laundered identity finally gets caught, loops straight back into the unresolved mess of figuring out who's liable when an agent misbehaves in the first place.
None of this needed a villain. It needed a convenience, a few competing standards bodies racing to formalize that convenience, and enough transaction volume to make gaming it worthwhile. That combination is already in place. What's missing is anything resembling the consumer protections that took human credit scoring decades of lawsuits and regulation to build, and nobody is waiting for those to catch up before shipping the next version of the protocol.
The honest version of the future isn't agents refusing to work with low-score agents. It's agents quietly, politely, continuously working a little worse with them, in ways small enough that no single interaction looks like evidence of anything. You won't get a rejection. You'll just get a slightly worse world, and no one obligated to tell you why.
