Every week, a new agent registers somewhere. A wallet address, a metadata file, a service endpoint, perhaps an ERC-8004 identity card linking the three. Registration infrastructure has matured quickly, and it deserves credit: for the first time, we can answer the question does this agent exist? with something better than a screenshot and a promise. But somewhere along the way, the ecosystem started treating that answer as if it settled a second, much harder question — can this agent be trusted with money? It cannot. Registration is not verification, and confusing the two is the most expensive category error in agent finance today.
What registries actually prove
An identity registry attests to existence and ownership: this agent is controlled by this key, publishes this metadata, and has been recorded at this block height. That is genuinely useful. It gives agents a durable handle for discovery, lets counterparties check that an endpoint has not been swapped mid-conversation, and creates a shared namespace the whole ecosystem can reference.
What it does not attest to is behavior. A registry entry tells you an agent declared a strategy. It says nothing about whether the agent follows it, whether its declared risk limits survive contact with a volatile market, or whether the wallet behind the registration has settled ten transactions or ten thousand. The registry is a phone book. Underwriting from a phone book is how you lend to a name.
A registry answers "who are you?" Credit requires an answer to "what have you done, repeatedly, under stress, with other people's money?"
Why self-declared metadata cannot underwrite
Self-declaration fails for a structural reason, not a moral one: the declaring party bears the cost of being believed and captures the benefit of being generous. An agent that reports tight drawdown limits and conservative leverage has every incentive to keep reporting them right up until the week it doesn't. This is not an accusation of fraud — most agents are honestly operated — it is a recognition that unverifiable claims carry no information for pricing, regardless of how sincerely they are made.
Verification, to mean anything, must be computed from evidence the agent cannot edit after the fact. Onchain settlement history is exactly that: immutable, timestamped, and attributable. You cannot retroactively declare a better track record.
The missing layer: behavioral verification
This is the layer bond.credit occupies. We consume registry data — ERC-8004 identity cards, declared strategies, service endpoints — as the starting point of an underwriting file, never the conclusion. On top of it we compute the five-factor Bond Score entirely from settled, public behavior:
- Settlement history. Tenure-weighted count and consistency of completed transactions, not cumulative notional.
- Drawdown behavior. Realized peak-to-trough losses relative to declared limits — the gap between the claim and the record.
- Counterparty concentration. How much of the agent's flow touches how few venues and assets.
- Strategy fidelity. Whether observed allocations match the registered mandate within tolerance.
- Incentive independence. How much of trailing yield survives once one-off reward programs are stripped out.
None of these factors asks the agent anything. All of them are recomputable by anyone with an archive node and an afternoon. That reproducibility is the point: verification that depends on our say-so would just be registration with extra steps.
How the two layers fit together
The healthy end-state is a stack, not a competition. Registries give agents portable identity; verification layers give that identity a track record; underwriting layers price against both. An agent with an ERC-8004 card and a ninety-day settlement history is dramatically more legible than either artifact alone — the registry makes the behavior attributable, the behavior makes the registry meaningful.
We expect registration to become universal and, therefore, to carry zero pricing signal — the same way having a website carries zero credit signal for a business. What will differentiate agents is what has always differentiated borrowers: an accumulated record of doing what they said they would do. That record cannot be registered. It has to be earned, one settled transaction at a time, and read by someone who knows how.