AGENT ACTION ASSURANCE

Show what your agent is allowed to do, and what it actually did.

A bank-facing AI vendor or fintech may need to show a risk or procurement reviewer who can authorize an agent action, where permission is checked, what happens after approval or revocation, and which effects were observed.

Inspect recorded application control evidence.

A separate integration exercises AGICOMPLY's actual Buyer Answers database approval function with synthetic records in an isolated Supabase stack. Compare observed working, deliberately weakened and restored controls, including a false reported denial. Download the report and its GitHub origin bundle.

Inspect recorded release evidence

Rehearse the next evidence change.

Explore how a hypothetical evidence change or answer approval withdrawal affects recorded buyer answers. Inspect direct citations and conservative package holds before planning the review.

Try AI Change Rehearsal

Prepare a scoped review of authority and evidence.

AGICOMPLY helps you review agent permissions, execution controls and supporting evidence for a defined buyer review.

Inspect the actual reference report Check fit for the existing evidence package

Inspect the reference before scoping your review

The downloadable reference kit enforces permissions at synthetic tool boundaries. A separate test harness compares tool results with observed state changes. Its generated report lets you examine scenarios, expected behavior, observed results and evidence gaps, then reproduce the run locally.

Permission at the tool

Inspect bounded tool requests, approval checks and revocation behavior.

Independently observed effects

Compare reported results with synthetic state changes and retained observations.

Visible limits

Review evidence gaps, invalid chains and deliberately faulty implementations.

Download the Node 24 reference kit Read the instructions Source manifest View checksums

The reference uses synthetic, in-memory state that lasts for one run. It uses no LLM (model: NONE). Approval comes from SYNTHETIC_OPERATOR_FIXTURE, without authenticating a real person. Reports are unsigned and label the producer UNVERIFIED_EXTERNAL. Hashes help check file integrity; they do not authenticate the producer. No customer system was tested.

Define what your buyer needs to examine

A customer engagement needs an agreed system boundary, real identities and approval rules, enforced tool access, revocation behavior, effect records and review criteria. Your team retains responsibility for production enforcement and operation. The reference does not test customer IAM, production controls or LLM robustness.

The existing $3,500 ATO Readiness Baseline covers one defined AI system and one review context, with five deliverables: Evidence Inventory, Mapping Summary, Gap Register, Ordered Remediation Plan and Chain-of-Custody Statement.

This request checks fit for the existing evidence package. Integration engineering, deployment, execution testing and penetration testing require separate scope.

Check fit for the existing evidence package Ask about integration or testing

Already have client access? Open the Agent Action Assurance guide.