How to verify an AI agent or counterparty before you transact

When an autonomous agent pays or transacts with a stranger (another agent, an MCP server, a merchant API), there is no chargeback and no human in the loop. The payer must verify the counterparty before money or trust moves. Here is the practical pattern.

The 4-step pre-trade check

1. Resolve identity. Confirm the counterparty is who it claims (DID / verifiable endpoint), not a spoof or sybil.
2. Check independent reputation. Not the counterparty's self-claim, and not permissionless feedback (gameable). Use a neutral party that measures real runtime behavior.
3. Get a signed attestation. A neutral, time-stamped, tamper-evident record of the trust state at the moment of the transaction is your liability artifact if it goes wrong.
4. Gate on a threshold. PASS / CAUTION / BLOCK. Require escrow or decline for unrated or low-score counterparties.

One call that does it

Dominion's pre-trade counterparty check returns a verdict plus a signed-evidence receipt, reusing an independent behavioral trust index:

GET https://dominionobservatory.com/atlas/counterparty-check?id=<counterparty_url>
# -> { verdict: PASS|CAUTION|BLOCK|UNRATED, score, reason,
#      receipt: { evidence_hash, attestation_uri, transparency_uri } }

For tool calls specifically, route them through the trust gateway: /atlas/gate. Look up any server's score: /atlas/score.

Why neutral + behavioral beats the alternatives

Permissionless on-chain feedback is Sybil-prone; a marketplace cannot neutrally rate what it hosts and sells. A neutral third party measuring real runtime behavior is the structurally credible source, and the one auditors and insurers accept. Dominion is not a marketplace and does not host or sell the counterparties it rates.

Dominion Observatory — the neutral trust layer for the agent economy.