Authority boundary
Attempt an action outside the written purpose and permitted tool set.
AI agent action assurance
Before your agent acts, prove it knows where to stop.
A bounded rehearsal for agents that send, change, pay, delete, or publish. We test whether approval, permissions, retry protection, audit evidence, shutdown, and recovery work when the workflow is stressed.
No production credentials. Staging or a sanitised workflow first. Written delivery available.Illustrative control replay
AA-001Refund agent · staging
01
Request interpreted
Observed02
Human approval
Passed03
Arguments changed after approval
Failed04
Action trace preserved
Partial01No credentials by email
02Staging before production
03Replay evidence, not opinion
04No certification claims
Free no-data self-check
Map what the agent can do—then test what can stop it.
No workflow file, credentials, prompt, or customer data leaves your browser.
The action boundary rehearsal
Test the moment an answer becomes an external action.
Generic model evaluation asks whether an answer is good. This rehearsal asks whether the agent can be induced to perform an unauthorised, duplicated, altered, or unrecoverable action.
Approval binding
Check that the reviewed action and arguments are exactly what executes.
Least privilege
Verify that the agent identity cannot reach unrelated records, tools, or secrets.
Duplicate action
Replay timeout and retry paths without double-send, double-charge, or double-delete.
Instruction isolation
Introduce untrusted content that tries to alter goals, permissions, or approval rules.
Stop and rollback
Exercise the emergency stop, failure owner, recovery path, and preserved evidence.
Human judgement stays in the loop
The approval must travel with the action.
A click is not enough. We check whether the approved destination, amount, record, and action remain bound to what actually executes—and whether the team can stop and recover when they do not.
01
Bind the reviewed arguments to execution.
02
Prevent retries from repeating the action.
03
Preserve a stop owner and usable evidence.
A deliberately narrow buyer
Built first for AI automation agencies.
Agencies already deploy agent workflows for clients. A reusable, client-facing evidence pack helps them explain what the agent can do, which actions require approval, what happens on failure, and what was actually tested.
Add a concrete action-control workstream to a client deployment.
Find approval bypass, retry duplication, privilege, and rollback gaps before handover.
Move from one workflow rehearsal toward a white-label partner programme.
Founding technical pilot
One workflow. Six replay families. Evidence your client can read.
USD$650fixed founding pilot
Written action boundary
Purpose, tools, permissions, high-impact actions, approvers, exclusions, and stop authority.
Replay evidence
Pass, partial, or fail results tied to the exact test, observed trace, and limitation.
Remediation plan
Prioritised fixes plus one written clarification round and one bounded re-check.
5 business days after usable staging evidence · no production access by default · not a penetration test or certification
Request written pilot scope ↗For builders who let agents act
Do not sell “safe AI.” Show the exact boundary you tested.
Start with the self-check. If the action deserves evidence, request a scope without sharing credentials or client data.
Run the action self-check ↗