Appearance
Why governed agents
The problem
Companies are giving AI agents access to finance, supply chain and customer systems. Most of that access is copied from a person, a service account or a vendor default. When an auditor asks why does this agent hold this permission?, the answer is usually a ticket comment or nothing.
Three facts make this hard:
- Access tools describe holdings, not need. A directory or identity governance (IGA) system shows what a person holds. It does not show which part of that access a given task needs.
- Runtime tools see behavior after the fact. They can show what an agent did. They cannot show whether the agent should have been able to do it.
- SoD rules were written for people. A mitigating control assigned to a person does not automatically cover an agent that person builds.
Copying a person's access, or a part of it, does not solve this. A person may hold privileged access their task never uses. A subset of their access is not, by itself, necessary or appropriate for an agent.
What DutyGraph adds
DutyGraph adds a record of the work itself. A request starts from Task Cards. DutyGraph joins them to the owner's holdings and the company's rules. It removes scopes the owner does not hold or the rules forbid, and escalates conflicts for a human decision. The people who already approve human access decide. DutyGraph records each decision.
Proving that each scope is the minimum the work needs is the target. Today it is Partial: the evaluator does not yet check requested scopes against what the Task Cards need. See The key join.
DutyGraph is designed to sit beside the tools a company already uses:
| Layer | Typical owner | What DutyGraph does with it |
|---|---|---|
| Directory and identity | IAM | Reads holdings from an uploaded export. Does not provision. Your IAM issues the agent identity. |
| Identity governance and SoD | Internal Controls | Loads rulesets and mitigating controls as versioned packs. Live import from an IGA product is Planned. |
| Access requests | IT service management | Names the request fields IAM needs. Sending a Decision Export is Planned. |
| Agent runtimes | AI platform team | Exports the approved agent "formatted for" a runtime. Can check calls routed through the DutyGraph gateway. |
DutyGraph decides and records. Your IAM provisions. See Approval, provisioning and revalidation.
The wider model
The governance join is one use of a larger asset: an AI-readable operating model of the company. DutyGraph maps how work actually happens, who owns it and which rules govern it.
The two halves need each other. Discovery gives governance the business context behind each permission. Governance gives discovery a practical reason to stay current, because action requires authority. Making that feedback run through automatic revalidation is Planned. See Three ways in.
INFO
Status. Built: the evaluator, CEIL-01, the SoD checks, the approval chain, the ledger and the gateway check. Partial: task-based justification of each scope. Simulated: identity and business systems in the sample companies. Planned: CEIL-02 enforcement, live identity and IGA connections, Decision Export and automatic revalidation.
Source: gitbook/product/governance/why-governed-agents.md in the DutyGraph repository.