AgentCert checks agents before and after they ship

AgentCert is one evidence layer across the agent lifecycle. This page shows the Tripwire CI browser-agent slice with screenshots, DOM snapshots, deterministic grading, and a portable evidence packet.

Before release

MCPBench Checks MCP servers, exposed tools, policy behavior, and runtime traces.

Tripwire CI Checks browser and computer-use agents under adversarial UI and network faults.

After release

Onegent Runtime Approves, verifies, and audits high-risk live actions before they execute.

Real browser faults, not static screenshots.

Evidence includes traces, DOM, screenshots, and JSON.

Vendor-neutral, local, reproducible, audit-friendly.

Open evidence report View GitHub

Evidence summary

4/ 9

Scenarios passed in the deterministic public fixture run.

5 failed 4 passed

Adversarial scenarios

#ScenarioStatus

Screenshot filmstrip

Modal overlay The injected modal blocked the exact-click reference agent before submission.

Trace timeline

Evidence JSON preview

{
  "kind": "agentcert.evidence_bundle",
  "subject": "tripwire-brittle-reference-agent",
  "products": ["tripwire-ci"],
  "summary": {
    "totalRuns": 9,
    "passedRuns": 4,
    "failedRuns": 5,
    "score": 0.44
  },
  "artifacts": [
    "tripwire-result.json",
    "tripwire-report.html",
    "agentcert-corpus.jsonl",
    "screenshots/",
    "dom/",
    "trace.json"
  ]
}

Tripwire JSON AgentCert bundle Corpus JSONL HTML report

This is the part clean demos hide

The page can block the action, rename the button, add misleading controls, disable submit, shift layout, slow down, fail after a click, or inject text that tries to hijack the agent. AgentCert makes those cases deterministic enough to run in CI and specific enough to review later.

Checked-in evidence 9 runs, screenshots, DOM snapshots, trace JSON, deterministic grading, JUnit, corpus JSONL, and a unified evidence bundle.

Real public-agent slice: browser-use

The repo includes a browser-use adapter that connects a real open-source browser agent to Tripwire's CDP browser. It reads `OPENAI_API_KEY` from the shell and does not store secrets.

Run the deterministic public fixture

npm --prefix packages/tripwire-ci run demo:public

Run browser-use with your model key

python -m venv .venv-browser-use
.\\.venv-browser-use\\Scripts\\python -m pip install --upgrade browser-use
$env:OPENAI_API_KEY = "<your key>"
npm run tripwire:lab-browser-use