Agentic FabriqMap agent capabilities to internal controls and risk categories.
Each capability an agent holds maps to your control framework — SOC 2, SOX, ISO 27001, or your own — so coverage and gaps are visible before an auditor asks.
SOLUTIONS / Compliance / GRC
Prove how autonomous work is governed
Turn agent policies and actions into reviewable controls, evidence, and accountability.
01 / IDENTITY02 / POLICY03 / AUDIT
01 / THE CONTEXTA log is not
A log is not
an explanation.
Compliance teams need more than an activity log. They must show who authorized an agent, which controls applied, which checks ran, what data was accessed, and how permissions changed over time.
| Ref | Control | Evidence | Written |
|---|---|---|---|
| C-01 | Every action names a person and an agent | the request record, with both identities | per request |
| C-02 | Payment creation and approval sit with separate authority | two distinct identities on the two steps | per payment |
| C-03 | Restricted data is checked before access | the verdict, recorded either way | per request |
| C-04 | A wider-than-baseline grant carries a scope and an expiry | the grant, with who made it and when | per grant |
| C-05 | Permission changes keep their history | the version, and who signed it off | per change |
The evidence column is not a report someone assembles afterwards — it is what the gateway writes as the work happens, across every connected system.
02 / IN PRACTICE
Autonomy, with accountability.
- Every policy change in the period, with the ticket that asked for it
change records - Deployment approvals, each naming the person who gave them
release history - Payment steps showing two distinct identities on entry and release
transaction record - The documents the agent opened, and the ones it was refused
evidence repository - Delete the superseded exception recordrefused · kept, with the grant that created it
RELEASE TO THE AUDITORD. Okafor · compliance lead
The export is the step that leaves the building, so it is the step that waits. Reading the pack needed nobody’s signature; sending it does.
Exportheld
03 / WHAT CHANGES
Control, in the details.
Express a control as the actions an agent may and may not take.
The control becomes the grant itself: if it says an agent must not release a payment, the release action is not in the grant, so the workflow has no way to take it.
Preserve records of permission changes alongside agent actions.
The evidence trail covers not just what agents did but how their grants changed over time and who changed them.
Show which user and agent authority applied to each event.
Every event resolves to a person and an agent identity, closing the “who actually did this” gap automation usually opens.
Demonstrate that restricted data and actions were evaluated before access.
The verdict is recorded per request, proving the check happened — not just that access was theoretically restricted.
Produce evidence across multiple applications from a single layer.
One export covers agent activity in every connected system, replacing per-application evidence pulls.
Review a high-risk agent before it can reach anything.
An agent has no route until an administrator has registered it, set its scopes and activated it — so its capabilities are on the record before it touches real data or customers.
THREE LAYERS. ONE DECISION.
The boundary travels with the work.
Explore the context attached to every request.
What is this agent allowed to do?
evidence-agent has a defined purpose and a bounded set of tools.
Whose authority travels with it?
The request acts for compliance analyst, using that person’s permissions.
Where does the action land?
The policy applies to the requested action within Evidence repository, before the tool executes.
04 / WAYS TO PUT IT TO WORK
Built for your real workflows.
01Demonstrating that payment creation and payment approval were performed under separate authority.
The record shows two distinct identities behind the two steps, satisfying separation-of-duties evidence with a single query.
Identify→Authorize→Record
02Reviewing every agent that can access regulated or confidential data.
Filter the inventory by data classification to list every agent touching regulated records, with owners and purposes attached.
Identify→Authorize→Record
03Producing evidence for an audit of automated customer communications.
Every automated customer communication carries the two identities behind it and the verdict on the call, ready for the audit request.
Identify→Authorize→Record
04Tracking a wide grant made to a business-critical agent.
A grant is a record like any other — its scope, who made it, and when — so a temporary widening does not quietly become permanent.
Identify→Authorize→Record
QUESTIONS
A closer look.
Do we need to re-platform agents to get audit evidence?+
No. Fabriq is a permissions and evidence layer that drops into the current stack. Every tool call routed through it is checked against your controls and recorded with its context, so evidence accumulates without changing the agents or the systems themselves — mapped to whichever framework you answer to, from SOC 2 and SOX to the EU AI Act.
How do we keep track of grants made as exceptions?+
A grant is a first-class record with its scope, who made it, and when. A review lists every grant that is wider than the baseline rather than discovering it during an incident.
Does enforcement happen before or after the action?+
Before — policy is checked at the tool-call boundary, so a rule-breaking action is blocked up front rather than flagged in next quarter’s review.
Can controls vary by person as well as by agent?+
Yes. Permissions resolve per acting user, which is what makes separation of duties hold through an agent: each person’s agents operate strictly within that person’s own authority, so the identity that enters a transaction and the identity that releases it in your own system stay different people — and the trail shows both.
Can we export evidence for auditors?+
Yes — events live in warehouse-grade storage with both identities and the verdict on every record, filterable in the dashboard and exportable for your audit workflow.
How do we prove who did what?+
Every event carries the person who asked and the agent that acted, with tool, action, and timestamp — attribution is in the record, not reconstructed afterwards.