Why your traces are not audit evidence
A trace is made for the team that runs the agent. It shows each model call and each tool call in order, with timings, so an engineer can find a slow step or a wrong answer. Evidence is something another party can check without taking your word for it. A trace is the first and not the second.
This guide says what a trace is for, why a trace store you control proves nothing to an outsider, what a signed record made from the same spans adds, and where that stops.
What a trace is for #
Tracing is built for debugging and monitoring, and it is good at both. It is also built to be changed. Traces are sampled to save cost, filtered, redacted, sent again after a failure and deleted when a retention period ends. Those are the right choices for an engineering tool.
Each of them is also a way for an entry to change or go missing without a sign.
A trace store you control #
The person asking is an auditor, a customer's security team or a regulator. They were not there when the agent ran. They have the trace and your account of it, and the trace cannot rule out three things.
- An edit. A span is a set of attributes your own process writes. Whoever holds the project's key or the database can write a span with any content, and it looks like any other span.
- A drop. Sampling drops spans by design. Nothing in a trace says how many tool calls there should be, so a complete trace and a thinned one look the same.
- A backdate. A span's start and end times are fields the sender sets. A span sent this morning can carry last week's date.
It makes no difference whether you run the trace store or a tracing service hosts it. The service keeps what your process sends, with the times your process gave, and cannot show what was never sent.
None of this says anyone did edit, drop or backdate anything. The trace cannot show they did not. The same holds for logs: Why your own logs are not evidence.
What a signed record adds #
One span processor, added where tracing is set up, writes a decision and an outcome record for each tool-call span. The trace stays as it is: the processor exports nothing, and your tracing tool receives the same spans as before. The set-up is on Audit log from OpenTelemetry traces, with a page for each tracing tool.
Each record then carries what a span does not:
- A signature and a chain. The record is signed with the agent's own key and holds the hash of the record before it. An edit breaks the signature; a removed record breaks the chain.
- An independent co-signature. A witness that is not you signs each batch of records and the newest record of each agent, so a record left out or a history rebuilt shows.
- Public time bounds. Each batch is written after a public randomness round and before a public timestamp, so its time does not rest on anyone's clock.
Signing and chaining happen on your machine, free and with no account. The co-signature and the time bounds come once the hashes are sent to a witness: the hosted service, or your own under Enterprise.
The limits #
- Before or after. A wrapped tool's decision is always written before the tool runs. A record made from a span is written when the span says it is a tool call. Pydantic AI and the Vercel AI SDK say so at the start, so the decision comes before the tool runs. The LangChain instrumentations of OpenInference and OpenLLMetry say so at the end, so it comes after. Every record says which it was.
- Sampling. A span the sampler drops never reaches the processor, so its tool call has no record. With an always-on sampler none is dropped: Sampling.
- Only what the span carries. The record holds the tool name, arguments and result that the instrumentation put in the span. A tool call with no span has no record.
- Protecting a tool needs the SDK. The processor never stops a tool. "Do not run the tool unless it is recorded" comes from wrapping the tool with the SDK.
- Evidence, not certification. A record proves who signed it, that it has not changed and when it existed. It never says whether the decision was right.
Where a limit matters for a tool, wrap that tool with the SDK: its decision is then written before it runs, and a protected tool is there when you want it.
Next: start free, with the SDK alone and no account. Quickstart