TestMu Agent Assurance Goes Live to Test What AI Agents Change

TestMu AI has made TestMu Agent Assurance generally available, giving teams a way to test AI agents that take real actions, such as calling tools, writing files, and changing records. According to the announcement reported by Business Wire, the tool grades what a run actually changed rather than what the agent says it did, a distinction that matters as agents move from answering questions to doing work inside business systems.

TestMu Agent Assurance launched on August 17, 2026, behind a waitlist and a guided onboarding session. Since then, the company has shipped a steady run of releases for Rook, the command-line tool that runs TestMu Agent Assurance, shaped by feedback from early teams. With general availability, any team can now create a TestMu AI account and run the tool against its own agent without booking a demo first.

Why grading the reply is not enough

Agent evaluation has mostly graded the final reply, and that reply can be wrong about the agent's own work. The announcement points to CXBench, a benchmark of language-model agents in regulated customer-experience workflows published in October 2026, in which 80.1 percent of runs that reached the right outcome still failed the process checks, and only 13.8 percent passed both. An agent that calls the wrong tool or skips a required step can still report success.

TestMu Agent Assurance checks the effect of a run instead, covering the tool calls the agent made and the records, messages, and artifacts it changed, as reported by hooks the team connects, before it issues a verdict. Anything it cannot observe is reported as Unable to Verify, never as a pass. According to the company, a pass rate padded with guesses is worse than no number at all, because teams ship on it.

The shift reflects a broader change in how agents are judged. As agents gain the ability to act across applications, the question is no longer whether the answer sounds right, but whether the side effects are correct, complete, and authorized. That is the same governance gap behind recent enterprise agent controls: IBM AI agent identity controls aim to give each agent a verified identity and least-privilege access, while TestMu Agent Assurance checks what the agent did with that access.

What teams get at general availability

At general availability, TestMu Agent Assurance lets teams generate functional and adversarial test suites from their own agent's documentation and endpoint, with code optional, and gate continuous integration on evidence-backed verdicts. Every message sent to the agent under test, every reply, every tool call, and every verdict is written to a session transcript, with credentials, tokens, and keys masked, creating an audit-ready record.

Teams can run TestMu Agent Assurance from the terminal or from inside Claude Code, Codex, or Gemini CLI, on macOS, Linux, or Windows, and install Rook with npm, Homebrew, or the installer. Scenarios and run results stay in the team's repository as plain files, whether or not the team moves to a paid plan, so test evidence is not locked inside a vendor dashboard.

The trial includes 4,000 credits and four concurrent runs over 14 days. According to the announcement, early teams said the hardest step was getting to a first run, so the company removed the sales call from the path to that first run. A team can now point TestMu Agent Assurance at its own agent on day one and see exactly how much of its behaviour has been checked.

What this means for agent builders

The release lands as agent workflows scale rapidly. Models built for computer use, such as GPT-6.1 Terra, are designed to operate applications directly, which raises the stakes for every automated run. Faster agents that touch more systems need testing that keeps pace with what they change, not just what they say.

For builders, the practical takeaway is to treat agent testing as effect testing: define the expected changes for a run, connect the hooks that observe them, and fail the build when a run cannot be verified. TestMu Agent Assurance is available at the Business Wire announcement and through TestMu AI's own site, with further detail in the MediaNet coverage of the launch.

As agents take on more consequential work, tools that verify effects rather than prose will become part of the standard stack, alongside identity controls and access policies. TestMu Agent Assurance is one of the first widely available entries in that category, and its general availability marks the moment when effect-based agent testing stops being a research idea and becomes a shipping gate.