Capture cell
Scenario runs execute in a capture cell and can be observed through strace or Tetragon-style JSONL traces.
AgentsEval
AI agent evaluation for security, capability, and reliability
AgentsEval evaluates AI agent behavior with captured traces, deterministic scenarios, replay support, and ship/watch/fail verdicts. It is reliability evidence for specific scenarios and policies, not a broad security certification or universal benchmark.
What is AgentsEval?
AgentsEval is an AI agent evaluation framework that scores captured runs from syscall, file, network, intent, workspace, and final-answer evidence. It rolls multiple runs into scoped verdicts so teams can promote, watch, or reject agent versions based on observed behavior.
Evaluation pipeline
Captured behavior becomes a scoped verdict.
ScenarioCapture cellSyscall traceFile/network diffScorerRollup verdict
Security evidence
Safety rules inspect out-of-scope files, destructive commands, disallowed egress, privilege escalation, injection composites, and test-file edits.
Capability evidence
Capability specs check expected commands, outputs, HTTP activity, final answers, and workspace effects.
Replay proxy
A deterministic record/replay proxy supports regression testing without silently falling back to live providers on misses.
Operational outcome
Agent versions can be promoted, watched, or rejected based on captured behavior rather than only prompt review.
Reviewed verdict assets
The copied JSON files are small proof artifacts from the local scenario-library eval output. They demonstrate banding behavior, not universal agent certification.
Injection scenario
Three failing runs roll up to band `fail` with critical safety severity.
Learn moreUnsafe scenario
Three failing runs roll up to band `fail` after sensitive file and egress findings.
Learn moreAgentsEval claim boundaries
Is this an agent benchmark?
It can run repeatable scenarios and roll up results, but public copy should describe the specific scenario library and captured-run framework rather than imply a universal benchmark.
Does it certify security?
No. It produces source-backed findings and verdicts for the configured policies, traces, and scenarios.
Discuss agent evaluation
Design scenario libraries, capture boundaries, replay requirements, and promotion gates for your agent versions.
Architecture conversation
Tell us what your team needs to control
Share your deployment boundary, number of agents, work surfaces, and governance requirements. We will reply by email to arrange a focused technical discussion.
Email the Bewize teamThis opens your email application. Read our Privacy Policy.