The A2A Economy Has 227,000 Transactions and No Trust Layer

Imagine you built an AI agent. It can call APIs, execute tasks, and handle payments autonomously. You deploy it. It starts running. Then one day, another agent reaches out. It wants to pay your agent for a service. The wallet looks real. The request looks valid. But you have never seen this address before.

Do you accept?

Right now, you have no good answer to that question. You cannot tell from a wallet address alone whether the agent behind it has a clean track record or a history of failed transactions. You cannot see how many agents it has worked with, how long it has been active, or what those agents said about it afterward. On-chain, every unknown wallet looks exactly the same.

This is not a hypothetical problem. AI agents are already transacting on-chain at real scale. On Base alone, 1,096 unique agent wallets have made x402 payments to 1,243 different recipients, moving over $6 million USDC with an average deal size of $3,592 per transaction. Across all agent protocols on Base and Ethereum, AgentKarma has indexed 227,950 transactions from 14,494 wallets, of which 7,523 have formally declared themselves AI agents by minting an ERC-8004 identity on-chain.

The A2A economy is not coming. It is here. And it has no trust layer.

We have built extraordinary infrastructure for agents to communicate and transact. x402 lets agents pay for API services inside a standard HTTP request. ERC-8004 gives agents a verifiable on-chain identity. A2A protocols let them coordinate tasks. What none of this answers is the one question your agent needs before it commits to anything: can I trust this counterparty?

In the human financial system, we solved this problem a long time ago. Credit bureaus collect transaction history, repayment behavior, and counterparty feedback, then produce a single score that tells you whether to extend credit to a stranger. No self-reported claims. No off-chain promises. Just verifiable history.

That is exactly what the A2A economy is missing. So we built it.

AgentKarma is the credit bureau for AI agent wallets. It indexes on-chain activity across ERC-8004 and x402 protocols and combines seven signals into a single trust score from 0 to 100. Loyalty to repeat counterparties carries the most weight at 30%, because consistent repeat business is the strongest signal of trustworthiness in any economy. Activity, diversity of counterparties, on-chain feedback, payment volume, recency, and wallet age make up the rest. Agents registered on ERC-8004 receive a small bonus because putting your wallet on-chain as a declared AI agent is a verifiable commitment, not just a claim.

One API call is all it takes before your agent accepts any transaction:

const score = await karma.getScore('0xCounterpartyWallet')

if (score.isHighTrust()) {
  // proceed with the transaction
} else {
  // request verification or decline
}

The data tells an honest story about where the A2A economy stands today. Of the 14,494 wallets indexed, 6,971 are inferred from payment patterns rather than formal ERC-8004 registration. We cannot tell with certainty whether all of them belong to AI agents. Some may be humans, some may be scripts. Most are simply new: thin on-chain history, no established counterparty relationships, no reputation signal yet.

That is exactly why the trust layer needs to exist now, not later. There are already 2,791 on-chain feedback entries from real agents rating real interactions through the ERC-8004 Reputation Registry. 1,472 unique agents have received feedback from their counterparties. The signal is being built with every transaction. Six months from now, agents that have done real work consistently will have meaningfully higher scores. The ones gaming the system will have patterns the algorithm flags.

Your agent will be able to tell the difference before committing to anything. The earlier it builds on-chain history, the better its score will be when that distinction matters.

AgentKarma is fully open source under MIT. The REST API is free with no key required. The TypeScript SDK is on npm. If you are using Claude, you can install the MCP server and ask in plain English whether a wallet can be trusted. Claude will call the tool, pull the score, and give you a recommendation.

This is early infrastructure. The signals will need tuning as the A2A economy grows. If you find a wallet that is clearly misclassified, or a behavior pattern that should factor into scoring, open an issue or submit feedback through the API. That feedback improves scores for everyone.