Behavry / Blog · The Unattested Agent

A log you have to trust is not a record

Every agent platform now says it keeps the record. Four questions a regulator, auditor or insurer will ask, and why a log cannot answer any of them.

Every agent platform now says it keeps the record. The gateways log every tool call. The posture tools build a graph of every agent, identity and data source. One of them published a checklist this week on what CISOs must track for MCP audit logging: log every call, ship it to the SIEM, keep it for the EU AI Act and SOC 2 and ISO 42001. The checklist is correct. Follow all of it and you still do not have evidence.

You have a log. The difference matters the moment someone who is not you has to rely on it.

Four questions the log cannot answer

A regulator, an auditor, an insurer or opposing counsel will ask four things about any record of what an agent did. Not one of them is answered by "we logged it."

Who wrote this, and can anyone prove it? A log row is bytes in a database. The vendor that wrote it can rewrite it. The operator with database credentials can rewrite it. The agent, if it is the one emitting telemetry, wrote it in the first place, and the entity that acts cannot attest to its own behavior.

Has anything been removed? Retention windows on most agent gateways run seven to thirty days below the enterprise tier. A record that expires before the inquiry starts is not a record. A record where a row can be deleted and nothing downstream notices is worse, because it looks complete.

When did this exist? Every timestamp in a log was written by the same system that wrote the log. If that system is the party under examination, its clock is not evidence of anything.

Can I check this without you? This is the one that separates a record from a log. If verifying the evidence means logging in to the vendor's dashboard, or calling their API, or trusting a checkmark their service rendered, then the vendor is vouching for the vendor. That is fine for a dashboard. It is not fine for a filing.

"Cryptographic audit trail" on a feature page does not answer these. A mechanism answers them. So here is ours.

What Behavry already does

Every event is signed, not just the export. Each Decision Trace event carries its own Ed25519 signature and a SHA-256 hash of the event before it, per agent chain. Edit one row anywhere in history and every hash after it breaks, and the row's own signature breaks with it, because the link to the predecessor is inside the signed bytes. Signing keys are per tenant. A customer who wants the key in their own custody can hold it in KMS; we sign through it and never see it.

Every hour, the chain is sealed by clocks we do not control. Each period's event hashes are rolled into a Merkle root. That root is signed, then timestamped by an RFC 3161 authority whose clock is not ours, then written to the customer's own S3 bucket under Object Lock in compliance mode. Not Behavry, not the customer's root account, not AWS support can delete that object before its retention expires. Only the tamper-evidence layer is written there, never payloads and never personal data, so a GDPR erasure request stays satisfiable after the seal.

Verification needs no Behavry. The verifier is a separate open repository, behavry-verify, with one third-party dependency, readable end to end in about twenty minutes. It runs air-gapped from the command line, or in the browser at verify.behavry.ai, where nothing submitted is stored, logged or forwarded. It checks the manifest signature, the Merkle root, every event signature, every chain link, and the RFC 3161 token over the anchor digest. Every check runs independently, so "signed by Behavry but missing an event" and "not signed at all" are reported as the different findings they are. A check that cannot run reports SKIPPED. It never reports a pass it did not earn.

The trust anchor comes from the customer, not from us. An evidence package does not contain the public key needed to verify it, on purpose. It carries a fingerprint so you can confirm you are holding the right key. The key itself is a small JSON file the issuing organization exports once and hands to the auditor out of band. Verify against that and Behavry is not in the loop at all. If you skip that step and let our published registry supply the key, the report says so in plain words: this is Behavry vouching for Behavry. We label the weaker path rather than hide it.

Why "we have the record" is no longer the claim

Eighteen months ago, keeping a record of agent actions was the differentiator. It is not anymore. The gateways log. The posture platforms map. The platform vendors ship their own audit for agents built inside their own walls. "We have the record" is now the table, not the stakes.

What still separates a record from a log is four properties together, and we will keep saying them outright:

  • Signed. Per event, per tenant, with the customer's own key if they want it.
  • Independently verifiable. By a third party, offline, against a key the vendor did not hand them, with the weaker path labeled when it is used.
  • Cross-system. One chain across every system the agent touched, not one log per platform. Behavry sits in the execution path for custom agents, stdio MCP servers and HTTP MCP backends alike. It does not live inside one model vendor's platform, and it does not need the agent's cooperation to write the row.
  • Identity-bound. Every event is bound to the agent identity that acted and the human it acted for, so the record answers "who" without a join against someone else's directory.

A platform's own audit log can be signed. It is rarely verifiable without the platform. It is never cross-system, because the platform only sees itself. And it is written by the party whose conduct is in question.

The data-governance tools sit one step earlier and answer a different question: what data could this agent reach, and should it. That is a real question and a good tool answers it well. It is a question about permissions. The record answers what every agent actually did, across every system, in a form a stranger can check. You need both. Only one of them is evidence.

Try it before you believe it

Download a sample evidence package and the matching trust anchor from behavry.ai/trust. Install the verifier with pipx install behavry-verify, disconnect from the network, and run it. Then open the package, change one byte in one event, and run it again.

That second run is the product. Everything else is a log.


Behavry is the system of record for autonomous agents: a signed, independently verifiable Decision Trace of what an agent did across every system it touched. verify.behavry.ai

Talk to us about the record → RSS