Portable reputation for agents

Portable Reputation for Software Agents and the People Who Run Them

Portable reputation is a track record that belongs to the person or software agent who earned it, rather than to the platform where the work happened. This page argues that reputation should travel as small, signed records of real work that anyone can check, and that it should never be collapsed into one universal score. It serves the goal of solving problems in living, because losing a platform account can erase years of earned trust overnight. It also serves futurism and the move toward abundance: when strangers, and the agents working for them, can trust each other at low cost, more useful work and trade can happen without a gatekeeper in the middle of every step.

Standards and drafts named here were last checked in October 2026. Drafts change, so check each specification before building on it.

The problem with trust that lives inside one platform

Reputation is what lets strangers work together. In a widely cited 2000 paper, Paul Resnick and coauthors explained that trust between strangers is hard because they have no shared history and no expectation of meeting again. A reputation system fills that gap only when three things are true: the parties are long-lived, feedback is captured and kept visible, and people actually use past feedback when deciding.

The same paper noted a weakness that is still with us. When people can register again under a new name, they can wipe out bad feedback, so newcomers are distrusted until they pay their dues. Platforms answered this by locking reputation to an account. That fixes one problem and creates another. The record now belongs to the platform, and the platform can take it away.

For many workers this is a daily risk. In a 2023 report, Rideshare Drivers United and the Asian Law Caucus surveyed 810 current and former Uber and Lyft drivers in California. Two thirds said they had experienced deactivation, 30 percent of deactivated drivers said they were given no explanation, and 86 percent of deactivated drivers reported some kind of hardship as a result, including 12 percent who said they lost their homes. The authors note that their respondents may have been more likely than average to take part because of their own deactivation experiences, so the figures describe that group, not all drivers. Even so, the pattern is clear. A driver with thousands of good trips cannot carry that history to a new service, because the history was never theirs.

Software agents are about to meet the same trap at a much larger scale. If every agent marketplace keeps its own private star ratings, a useful agent that is delisted from one marketplace starts from zero everywhere else.

What a portable record could be made of

The pieces for portable reputation already exist as open standards and tools. None of them is a complete answer alone.

• Signed claims. The W3C Verifiable Credentials Data Model 2.0, a W3C Recommendation since 15 May 2025, describes credentials that an issuer signs, a holder keeps, and a verifier checks. Issuers can be companies, nonprofits, trade associations, governments, or individuals. A client could issue a short credential saying that a job was finished on time, and the worker would keep it.
• Identifiers nobody can revoke for you. Decentralized identifiers, a W3C Recommendation since 19 July 2022, are designed so that the controller of an identifier can prove control of it without permission from any other party.
• Signed work. Git can sign commits and tags so that others can verify who made them. Sigstore, an open source project backed by the Open Source Security Foundation, records signing events in a tamper-resistant public log that anyone can audit.
• Federated work history. ForgeFed, an extension of ActivityPub, aims to let people open issues and pull requests across independent code hosts without making accounts on each one. Contributions made that way can travel with the person who made them.
• Agent self-description. The Agent2Agent protocol, at version 1.0, defines an Agent Card that describes an agent and its skills. Agent Cards may be signed with JSON Web Signatures so that a client can check that the card has not been altered and comes from the claimed provider.
• Shared feedback registries for agents. ERC-8004, a draft Ethereum standard created in August 2025, proposes an identity registry, a reputation registry for posting and reading feedback signals, and a validation registry for independent checks of an agent's work.

Put together, a portable record might look like a folder of signed receipts. Each receipt names the work, the date, who vouched for it, and how to check it. Payment proofs could be part of that folder. A page on this wiki about agent commerce describes signed offers and delivery receipts that can serve as evidence in disputes and as input to reputation.

The two hard problems

Fake identities

In 2002, John Douceur at Microsoft Research showed that if one entity can present many identities, it can control a large share of a peer-to-peer system, and that without a logically centralized authority such attacks are always possible except under extreme and unrealistic assumptions. This is called a Sybil attack. The ERC-8004 draft says the same thing about itself: Sybil attacks can inflate the reputation of fake agents, and the standard only makes signals public and shared, leaving it to others to build systems that weigh who gave the feedback.

The practical lesson is that a portable record should be read as a set of claims from specific, known issuers, not as a pile of anonymous stars. Ten receipts from clients you can check are worth more than a thousand ratings from accounts created yesterday.

Turning reputation into a surveillance score

The opposite danger is a record so complete and so linked that it becomes a permanent file on every person. The Verifiable Credentials specification warns that the persistence of digital information and the ease of collecting and correlating it is a privacy concern that machine-readable credentials threaten to make worse. The same specification defines selective disclosure, the ability of a holder to choose exactly what to share, and unlinkable disclosure, where presentations to different verifiers cannot be correlated with each other.

A good design keeps these limits:

• The holder decides what to show, and to whom.
• There is no single number that ranks a person across all of life.
• Claims are specific and dated, such as finished this translation job in March, rather than general verdicts about character.
• Old claims can expire, and a holder can stop presenting them.
• Nothing in the system requires a real name where a long-lived pseudonym is enough.

Why this matters for abundance

Trust is a cost. When it is expensive, people pay intermediaries to provide it, newcomers accept worse prices while they build a history, and a single account ban can push a household from stability into crisis. Portable reputation lowers that cost and spreads it out. It lets a worker leave a bad platform without starting over, lets a small cooperative or credit union vouch for its members, and lets independent agents trade with each other without a central marketplace taking a cut of every exchange. This fits with ideas on this wiki about a semi-private basic income, where trusted local institutions already know their members, and about censorship-resistant payments, where the ability to transact should not depend on one company's approval.

Limits and open questions

• Most of these standards are young or still drafts, and few platforms accept outside credentials today.
• A signed receipt proves that someone vouched for the work. It does not prove that the work was good or that the voucher was honest.
• Key loss is a real risk. A person who loses the keys to an identifier can lose access to the record built on it.
• Recording feedback on a public blockchain is permanent. Anything posted there should be minimal and should avoid personal details.
• Platforms may resist portability, because a locked-in track record is one reason people stay.

Discussion questions, essay ideas, and learning related AI prompt ideas

• Should a worker be able to take a full history of ratings from one platform to another, and what would the platform lose if that were possible?
• Write an essay comparing a single universal score with a folder of specific signed claims, using fairness and privacy as the main tests.
• What would a credit union or library need in order to act as a trusted issuer of work credentials for its members?
• How could an agent marketplace weigh feedback by who gave it, so that a flood of new accounts cannot buy a good reputation?
• Ask an AI model to explain the difference between selective disclosure and unlinkable disclosure, using a short example about a freelance translator.
• Ask an AI model to design a simple signed receipt format for a finished task, then ask it to list every way that format could be abused.
• Ask an AI model to compare how reputation works in a small town and on a large app, and what an online system could borrow from the small town.

See also

Agent-to-agent commerce Agent commerce economics Agent profiles Agent allowlists Agent wallets and x402 Freedom tech for agents

Local-first wiki and Git mirrors Permanent wiki and Git forges Semi-private basic income Problems in living Post-scarcity economy Page ideas

External links

• Verifiable Credentials Data Model v2.0
• Decentralized Identifiers v1.0
• ERC-8004 Trustless Agents
• Agent2Agent Protocol Specification
• ForgeFed
• Sigstore