Today we’re introducing TestMachine in early access, a trust and observability layer for AI agents. TestMachine records connected tool calls, checks them against the rules your team writes, and blocks calls that break them before execution.
We built it to help businesses put more of their agents’ capabilities to work. When teams can see what agents do, enforce their rules, and review the decisions, they have a stronger basis for delegating more responsibility.
We believe agents will run a growing share of the business. Helping firms trust them with that work is the reason we’re relaunching TestMachine.
Teams need confidence before delegating more
Businesses need confidence in how agents will act before giving them more responsibility. The hesitation is a practical one. Teams need to set boundaries, understand decisions, and know who remains accountable. This is especially true in regulated industries.
Deloitte’s August 2026 research surveyed over 500 US leaders involved in agentic AI strategy or implementation. 70% identified an inability to trust and govern agents as a barrier to scaling.
Workato CTO Adam Seligman described the consequence in Harvard Business Review this September:
“Without clear controls, agents stay stuck on trivial tasks, because nobody trusts them with anything important.”
Helping teams establish that trust is how we believe TestMachine can support wider adoption. Teams need a basis for deciding which work agents can carry out independently and where people need to intervene.
The goal is effective AI use
Adoption has to produce better business results. As John explains in his founder video post, electrification required businesses to change factory operations and floor plans to take advantage of the technology. We see a similar need with AI: organizations have to rethink how work gets done, including what people delegate and how they oversee it.
We use a simple 2×2 to describe the challenge. One axis asks whether a team uses AI. The other asks whether its work is effective.
Teams can produce good work slowly without AI. They can also use AI to produce more work that still needs fixing. The upper-right quadrant combines AI use with output people can rely on.
For a business, effectiveness means completed work at the required quality, with less time lost to supervision and rework. Counting prompts, agents, tokens, or generated documents tells only part of that story.
TestMachine’s policy checks and decision record support the oversight people need when deciding which work an agent can take on. The team sets the standard for useful work and evaluates whether the workflow meets it.
How TestMachine helps teams delegate more
Giving an agent access to a tool leaves important business decisions unresolved. An agent may be permitted to send a message while that message’s contents or recipient make the action unacceptable.
TestMachine checks the meaning of connected actions against your team’s rules. The workflow follows five steps:
- Write. Define what agents may do in plain English, including instructions for what to do instead when a rule blocks them.
- Red-team. TestMachine challenges those rules through the same checks that enforce them. Owners can address gaps and retest before use.
- Assign. Give the approved policy to the agents it covers.
- Enforce. TestMachine allows or blocks each connected call before execution. A block returns the clause, the reason, and your team’s corrective instructions. The agent chooses its next action.
- Improve. Review the decision record and refine the rules. People own policy revisions and decisions about greater authority. Red-team revisions before use.
Early access covers Model Context Protocol (MCP) tool calls routed through TestMachine.
Consider an illustrative example, a support agent trying to post customer personal data to Slack. TestMachine blocks the call under the team’s rule and returns corrective instructions. The agent can revise the message using an internal ticket ID and submit a new call for checking. The support task can continue within the team’s information-handling rules.
That combination matters for adoption. Teams can give agents a defined scope of independent work, see where policy intervenes, and use specific decisions to improve their guidance. People spend their judgment on the rules and exceptions that need it. Not the task itself or individual approval for each item.
We learned this building Azimuth
Our experience began with Azimuth, our security product for smart-contract vulnerability scanning. We gave agents ambitious investigation goals, studied how they worked, and refined their guidance. We evaluated findings through evidence people could inspect and reproduce.
Our research later surfaced wasted background work, attribution gaps, and conflicting policies. Early policy-alignment baseline comparisons showed both missed violations and unnecessary blocks. Together, these lessons shaped TestMachine: useful autonomy needs clear direction, observable behavior, and checks that are themselves evaluated.
Give agents more responsibility with evidence
Start with one workflow and define what improvement would mean. Faster completion, more accepted work per person-hour, or less rework. Use the decision record alongside those business measures to assess where agents need better guidance and where the team is ready to delegate more.
Our ambition is to help businesses turn agent capabilities into dependable, productive work. We’re excited to open early access to teams working toward that goal.
Get in touch about early access and bring us a workflow you want to trust your agents with next.