AI Agents Tested on Actions, Not Just Words

TestMu AI has announced the general availability of Agent Assurance, a novel Quality Engineering platform that specifically tests AI agents by focusing on their actual actions rather than just their reported outcomes. Launched on August 17, 2026, this system evaluates what an agent has changed, including tool calls and data modifications, ensuring a more accurate assessment of its performance. Standard agent evaluation often overlooks process failures, with many runs passing despite incorrect steps. Agent Assurance addresses this by verifying tangible effects and reporting 'Unable to Verify' for unobserved actions, never counting them as passes.

The platform offers auto-generated test suites from agent documentation and endpoints, with optional code integration for deeper insights. It includes advanced red teaming features with 25 attack patterns mapped to OWASP Top 10 risks and MITRE ATLAS techniques. Evidence-backed verdicts are provided, clearly stating Pass, Fail, or Unable to Verify, supported by concrete evidence. Beyond functional tests, it assesses quality, performance, token economy, and latency, with root-cause analysis available upon request.

Agent Assurance facilitates regression testing in CI/CD pipelines, supports multi-agent testing, and allows controlled writes with user confirmation. It operates without provider keys, running models in TestMu AI's controller, and offers an audit-ready record of all interactions. The tool is accessible via terminal or IDEs on macOS, Linux, and Windows. A 14-day free trial is available, providing 4,000 credits and four concurrent runs, with scenarios and results stored as plain files.