A score predicts; a bond pays

In January 2026, ERC-8004, the Trustless Agents standard, went live on Ethereum mainnet: a deployed standard, not a pitch deck, with contributors from MetaMask, the Ethereum Foundation, Google, and Coinbase. For the first time, an autonomous agent can carry a portable on-chain identity and a public reputation.

That matters because once software starts transacting on its own, the first question is the oldest one in commerce: can I trust the counterparty? ERC-8004 gives that question an answer you can read on-chain.

But knowing who to trust is not the same as being protected when the trusted party fails.

Nobody has standardized that second half. That’s the half we build.

What a reputation score really says#

ERC-8004 organizes agent trust into three registries. Identity says who the agent is: an on-chain record minted as an ERC-721 “agent card,” pointing to its endpoints and a payment address. Reputation is a track record others can score, deliberately modeled on a credit score. Validation lets an agent prove it ran what it claimed without revealing the model behind it.

It is good architecture. But notice what reputation is: a prediction. A score compresses an agent’s past into a probability about its next move. Credit scores do the same for borrowers. Useful, often necessary, and sometimes wrong.

What scores cannot repay#

Here is what a score cannot do. When the well-rated agent fails on your transaction, shipping the wrong thing, misreading the intent, or going dark mid-task, the score updates. For the next counterparty. You are still out the money.

Reputation disciplines the population over time. It does not make the individual harmed party whole. A credit score never refunds the lender it just misjudged; it adjusts so the next loan is priced a little better. The loss already landed on someone.

For human commerce we paper over that gap with institutions that actually pay: the chargeback, the escrow agent, the insurer. For agents transacting at machine speed and machine volume, “your score will go down” is not recourse. It’s a footnote on a loss you have already taken.

Recourse is the missing half#

The primitive is simple to state: a bond posted before the agent acts, and a verified failure that pays the harmed party back, automatically and in private.

Reputation tells you who to trust. Recourse makes them pay.

This is the category we have been building: Bonded Settlement: settlement that carries its own recourse. An actor posts a bond before it acts; a verified failure pays the counterparty back. Put that under an agent and its promise stops being a claim and becomes collateral.

Reputation needs recourse#

They are not rivals. They are two halves of one mechanism, and each is weaker alone.

• Reputation without recourse is advisory. A rating with nothing standing behind it is a guess you are free to act on at your own risk.
• Recourse without reputation is mispriced. Every agent posts the same flat bond, the careful and the reckless alike, and the honest ones subsidize the rest.

Together they close the loop: reputation prices the bond, the bond settles the failure, and the settlement feeds back into the reputation. A trusted agent posts less and moves faster; a failing one pays the party it harmed and is repriced for everyone after. Accountability runs both ways. A track record you can price and bond against stops being a label and starts being an asset.

Built in, not bolted on#

There is a tempting shortcut: bolt recourse on afterward, with an insurer you call when the agent misbehaves, a dispute desk, or a manual clawback. We made the case against that in the human commerce post. Recourse bought separately is slow, intermediated, and hands your book to whoever underwrites it. For agents it is worse, because the speed and volume that make agent commerce interesting are exactly what a manual dispute process cannot keep up with.

So the bond is not a product beside the rail. It is part of settlement. Three primitives, one flow:

• Live Escrow. Funds held in confidence until the outcome is known, with no party holding them in the middle.
• Live Gate. Release on a verified event, not a human sign-off — a pluggable verifier behind the IConditionResolver interface.
• Testnet Recourse. Coverage that pays out when a counterparty fails. The contract flow runs with test assets; real risk pricing and a capitalized pool are next.

ERC-8004 reputation is a natural input to that pricing. The same score that tells you who to trust can help set what their bond should cost: a low-risk agent posts less, a shaky one posts more. We are not there yet; real risk pricing is the round’s work. But an open reputation registry and a priced bond are meant to compose.

The registries are public. Commerce is not.#

There is another reason recourse has to live in the rail. ERC-8004’s registries are public by design, and the payment standards agents settle on are open too. That transparency is a feature for discovery and a liability for commerce: an agent’s spend, its counterparties, its exposures, all legible to every competitor watching the chain.

Real businesses will not run their book on a billboard. Bonded Settlement is confidential by default, FHE-encrypted via Fhenix CoFHE: the bond and payout are real, while the amounts and counterparties stay private. The design goal is to price recourse over encrypted data without ever publishing the underlying book. Privacy here is a requirement, not an extra.

Where this fits#

None of this replaces the standards. ERC-8004 answers identity and reputation. Payment protocols like x402 answer authorization: may this agent pay? Agent-to-agent and tool protocols answer how agents talk and act. Bonded Settlement answers what happens after: where the funds wait, what proof releases them, and what recourse follows failure.

The Reineira Settlement Standard (RSS) is the open, forkable specification of those settlement primitives, and we are building an x402 adapter so the recourse layer plugs in where agents already pay. The protocol fee is fixed at zero; plugins may set their own. We are not standing up a competing rail. We are filling the layer the others leave open.

What is live now#

Live on Arbitrum Sepolia: confidential escrow and reference Gates. The recourse contract flow runs with test assets. Next: real risk pricing and a capitalized pool. There is no mainnet deployment and no production pool yet.

One worked example is Prova, a pre-launch trade-credit application built on these primitives. It demonstrates the intended flow on testnet for human counterparties. Production underwriting and capital are not live. An agent is the same primitive with a different actor posting the bond.

If you are building with agents#

Reputation is finally becoming real infrastructure for agents, and that is the right direction. But a score is a prediction, and a prediction is not a promise. The promise needs collateral. So ask it this way: once your agent’s reputation can carry a bond that actually pays when it fails, what would you let it do that you would not today? Three places to go next:

• Whitepaper (PDF) — the formal specification of the primitives and the standard.
• Quickstart, to build your first testnet flow.
• Telegram — tell us what your agent should be allowed to do once it can post a bond. We read every message.