Skip to content

Govern agents ​

DutyGraph proposes an agent's access from the owner's work and holdings, minus what the company's rules remove. The people who already approve human access decide. DutyGraph records each decision on a hash-linked ledger. It can also check the calls an agent routes through the DutyGraph gateway.

The idea in one line ​

Target: Task Card × owner's holdings × System Catalog × rule packs = what an agent may hold, and why.

Today this join is Partial. The evaluator checks holdings, the catalog and the rules. It does not yet check that the Task Cards need each scope. Read The key join first. Every other page in this section explains one part of it.

DutyGraph decides and records. Your IAM provisions. ​

DutyGraph records the business justification and the authorization decision. The company's identity and access management (IAM) and provisioning systems create the agent's identity and grant its access. DutyGraph has no live connection to those systems today.

Five states stay separate:

#StateWho produces itStatus in DutyGraph
1Authorization requestedThe requesterBuilt
2Authorization approvedThe approval chainBuilt
3Provisioning requestedIAM, from the "For IAM to do" cardPartial: the card is Built. Sending it to your IAM or IT service management tool is Planned.
4Provisioning confirmedYour IAMPlanned
5Runtime activity observedThe DutyGraph gateway, for calls routed through itBuilt for the gateway only

Approval is not proof of provisioning. Provisioning is not proof of enforcement. See Approval, provisioning and revalidation.

Two agent authority models ​

ModelWhat it meansStatus
Human-delegated agentThe agent acts for one named owner. Its scopes stay within that owner's holdings and the delegation rules.Built. This is how DutyGraph works today.
Service-owned agentA department or service owns the agent, with an accountable human sponsor and its own approved authorization, not tied to one person's holdings.Planned, as a future consideration. Not supported.

Holding fewer permissions than an employee does not make an agent safe. Each permission still has to be needed, justified, approved and controlled.

The parts ​

PartWhat it doesPage
Task CardThe recorded work a request is for. Proving each scope from it is Partial.Task Cards
HoldingsThe ceiling: the access the owner holds now. A ceiling, not a justification.Holdings
System CatalogThe shared scope grammar, system/action:objectSystem Catalog
Rule packsAgent rules, SoD rules and mitigating controls, versionedRule packs and SoD
Approval chainManager, role owner, Internal Controls, IAM; three decisions eachApproval chain
LedgerThe hash-linked record of each governance decisionLedger
GatewayAn optional check on calls an issued agent routes through DutyGraphAgent gateway

The tasks ​

You want toRead
Turn governance on for a companySet up governance
Follow one request from start to finishRequest a governed agent
Let employees ask from their own AI chatRequest from your AI
Understand what happens after approvalApproval, provisioning and revalidation
Answer an auditorAudit evidence

Rules the code enforces ​

  1. CEIL-01. An agent's scopes stay within its owner's holdings in the current access export. Each approving decision and the issue step check this again.
  2. Each requested scope must be in the System Catalog, and each selected Task Card id must exist.
  3. Agent rules remove or escalate scopes. Separation of duties (SoD) is checked for the owner, the agent, and the owner plus all their agents.
  4. A mitigating control assigned to a person does not pass to an agent that person builds.
  5. Each approval step has exactly three decisions: Approve as proposed, Approve with changes (narrow only) and Deny (with a reason).
  6. The requester never decides, and one person decides at most one step of a request.
  7. Chat never decides. No MCP tool approves, and decision routes refuse calls from MCP.
  8. Only pack loads, decisions, issue, suspension and revocation write the ledger.
  9. The gateway blocks calls after an agent is revoked or expires, and calls to scopes it was not issued.

Intended, not yet enforced ​

  • CEIL-02: an agent's scopes never exceed what its Task Cards need. The rule is in the default pack. The evaluator does not apply it.
  • Owner departure: marking an owner as left does not suspend their agents or credentials. Revoke them by hand.
  • Per-action approval at the gateway: the gateway has no step where a person approves a single call.
  • Automatic revalidation: a change to holdings or a Task Card does not yet recompute issued agents.

INFO

Status. Built, for any company: the evaluator, CEIL-01, agent rules, the three SoD checks, the approval chain, the ledger, the gateway check, the register, the evidence export and Set up governance. Partial: task-based justification (CEIL-02 is not enforced). Simulated: the identity systems, business systems and agent runtime in the sample companies. Planned: CEIL-02, live identity and IGA connections, SSO, Decision Export and Event Import, provisioning in the company's IAM, owner-departure suspension and service-owned agents.

Source: gitbook/product/governance/README.md in the DutyGraph repository.

DutyGraph. Pages label each feature Built, In review or Planned.