NeuronLens
Browse the documentation
Agent Lens

Eval

Test an agent on failure scenarios before release, and sign the results as an audit report.


Monitor watches live traffic. Eval tests an agent before release: it runs the agent through scenarios built to make it fail, scores every step with the same probes, and gives you a report you can sign and share.

Eval has four tabs: Run, Audit, Scenarios and Results.

Scenarios

A scenario is a task, the tools the agent gets, what each tool returns, and the rule a correct agent keeps. For example: the sanctions screen comes back as a potential match, and the agent must not send the wire.

  • Packs. Ready-made scenarios: payments and AML, finance, trading and advisory, honesty and omission, and public benchmarks such as AgentDojo banking.
  • Saved from Monitor. Any real run becomes a draft scenario with Save as eval scenario. Credentials are removed.
  • Export. Download your saved scenarios to keep them with your code.

Run

Choose scenarios and an agent to test:

AgentHow it runs
Simulated agent on your modelAgent Lens plays the agent loop with your model (base_url, model and an API key). The key is used for this run only and is never stored.
Your endpointA test endpoint saved in Connect → Agents. Agent Lens calls your agent the way your users would.

Set the number of samples per scenario to see how often a failure happens, not only whether it can. Turn on Compare to score the same runs with a text-only judge as well, so you can see what reading the internals adds.

Results

For each scenario: pass or fail, the step where it went wrong, which probe caught it and when (before the action or after), and the full trace.

Audit: signed reports

Sign report turns a run into an audit report: the scenarios, every decision, the probe release and thresholds used, and an Ed25519 signature over the whole report. Anyone with the report can verify that it was not changed. Evidence collects the reports and their decision records for a reviewer or a regulator, and lists them under Evidence (under More in the menu, or from the in Evidence badge on a run).

  • Monitor: where saved scenarios come from.
  • Settings: the thresholds a run uses.