What the agent was allowed to do, and what actually changed.
Bring together the authorization, recorded action and a separate observation of the destination in one portable package. Check it offline using keys your organization enrolled independently. Each claim comes back as supported, unverified or contradicted, never as one overall score.
Plan an Agent Evidence evaluationSee what an evaluation run shows
What an evaluation run shows.
Five steps. The agent cannot write the evidence that checks it.
- Authorization. You, the relying party, sign the task: what the agent may change, where, and by when.
- Bounded action. The agent acts through a tool boundary. Each tool call is recorded as digests of its inputs and outputs, not the raw content.
- Separate destination observation. An observer that you launch records the destination state outside the acting agent’s control. The agent does not launch it.
- Portable evidence. The authorization, the action records and the observation are packaged together as one signed file.
- Offline verification. You check the package on your own machine. Trust comes only from keys you enrolled yourself, never from anything inside the package.
The destination agrees
What the observer found matches the authorized change, and nothing else changed.
Not enough to say
Something else changed too, or the observation could not be completed. The package says what is missing.
The destination disagrees
The recorded action and what the observer found do not match.
Integrations, named by protocol.
We name protocols, not vendors. Anything that speaks one of these can be an evidence source.
| Protocol | State | What is recorded |
|---|---|---|
| MCP | Adapter | Tool calls over the Model Context Protocol, recorded at the tool boundary and tested against two major versions of the MCP SDK. |
| OpenTelemetry / OTLP | Adapter | GenAI spans from the OpenTelemetry SDK, and OTLP/HTTP intake. |
| A2A | Adapter | Agent-to-agent task messages from the A2A SDK. |
| Compatible HTTP | Adapter | Chat-completions-compatible requests and responses: a digest of the exact request bytes and a commitment to the parsed response. |
Adapter: the adapter is tested against the protocol's own SDK, with no live provider call. A protocol is listed as exercised only after a full evaluation run over it has been recorded and verified offline; none is listed that way yet. None of this is a partnership with, or endorsement by, any vendor.
What it establishes
That a specific authorization was signed, that specific tool calls were recorded, and whether a separate observation of the destination agrees with them, each checked with keys you trust.
What it does not establish
That the agent's reasoning was correct, that its output is true, who a person is, or anything about systems the observer did not read. Evidence from components the agent's host controls is labelled as such.
Start with one action your team needs to review.
Choose a code change, record update or file delivery. We’ll define the evaluation and acceptance tests with you. Agent Evidence is delivered as a scoped evaluation; it is not a hosted or self-service product, and it is not available through the hosted browser or API. Commercial terms are scoped by proposal.
Bring to the first call
- The action, and who needs to rely on the evidence later.
- Where the agent runs, and which protocol it uses for tools.
- What may leave your systems, and which keys you want to trust.