Appearance
FAQ
Each answer says what is Built, Simulated or Planned today, and links to the page with the detail.
For buyers and leaders
What does DutyGraph do today?
It maps how a company's work happens, as Task Cards in a Company Repo. It finds where an agent could help. For a requested agent, it proposes scopes from the owner's Task Cards and access, removes what the owner does not hold or the company's rules forbid, and escalates conflicts. The company's own approvers decide, and every decision goes to a hash-linked ledger. Checking each scope against what the Task Cards need is Partial today. See What is DutyGraph.
Does DutyGraph run my agents?
No. DutyGraph decides and records. Your IAM team provisions the agent, and the agent runs in your own runtime. DutyGraph exports the approved agent formatted for that runtime. An issued agent can call its tools through the DutyGraph gateway, which checks every call routed through it. Planned: a company can opt in to have DutyGraph create the approved agent on the company's own vendor account. The vendor still runs it. See Exports and runtimes.
What is simulated in the demo?
In the sample companies, the identity systems (Okta, Entra ID, Saviynt, ServiceNow), the business systems that answer agent calls, provisioning and the agent runtime are Simulated. The people, access, hours and predictions are illustrative. Approvers are sample personas that the advisor operates, and each decision records "operated by" that advisor. The evaluator, approval chain, ledger, gateway check, MCP request tools, register and export are Built. Cobalt Industrial Supply (fictional sample) sends no messages.
Does DutyGraph hold certifications or have proven customer results?
These docs make no claim of certification, customers or results. Built means the code exists and its tests pass. It does not mean a customer runs it in production. See Built and planned.
For IAM and Internal Controls
Does it connect to Okta today?
No. Today DutyGraph reads an access-export CSV that you upload in Agents › Rules › Set up governance. Read-only connections to Okta, Entra ID and Saviynt, set once per company, are Planned. DutyGraph never collects employees' own logins to business systems. See Holdings.
Can an agent get more access than its owner?
No. Rule CEIL-01 keeps an agent's scopes within its owner's holdings, and rules only remove or escalate a scope.
Being within the owner's holdings is a ceiling, not proof of need. An owner may hold access their task never uses. Today the approval chain judges need. The evaluator does not yet check scopes against what the Task Cards need (rule CEIL-02 is Planned). See The key join.
Who approves an agent request?
The people who already approve human access. The default approval chain is manager, role owner, Internal Controls, then IAM. Each step has three decisions: Approve as proposed, Approve with changes (narrow only) or Deny (with a reason). Only the current step's assignee can decide, and one person can decide only one step of a request.
Can the employee approve their own request?
No. The requester's view has no decision control. The server refuses a decision from the requester, even if they are named on a step.
Can an AI chat approve or change a request?
No. No MCP tool can approve, change or deny a request, and decision routes refuse calls that come from MCP. Approval happens in DutyGraph. See Request from your AI.
Does a person's SoD mitigation cover the agent they build?
No. A mitigating control covers one principal and one risk. If the agent needs conflicting scopes, Internal Controls must tag a control for the agent itself, and only one marked as covering agents. DutyGraph also checks the owner plus all their agents as one set. See Rule packs and SoD.
Does DutyGraph provision the agent's identity?
No. DutyGraph decides and records. Your IAM provisions. At issue, DutyGraph shows a "For IAM to do" card that names the agent identity to create, with exactly the approved scopes. Sending the decision to your IAM or IT service management tool (Decision Export), and reading back what was provisioned (Event Import), are Planned. See Approval, provisioning and revalidation.
What happens when access or rules change?
A new rule or holdings file is a new pack version, and the old one stays on record. A pending request's next decision is refused as stale until the reviewer reloads. A saved builder manifest turns Broken when a role holder leaves or the recorded access or rules become stricter. An issued agent is not recomputed: it keeps its issued scopes until it expires or is revoked. IAM or Internal Controls can revoke it, and the gateway blocks calls after its expiry. Automatic revalidation, recertification and drift checks are Planned.
What happens to an agent when its owner leaves?
Today, nothing automatic. Marking a person as left ends their private links and disconnects their phones. It does not suspend or revoke their issued agents or the agents' credentials. Revoke each agent by hand. The person can no longer use the MCP request tools, and a saved builder manifest that names them turns Broken. Automatic suspension on owner departure is Planned.
Can a person approve each action an agent takes?
Not through DutyGraph. The DutyGraph gateway has no step where a person approves a single call. Where a runtime has its own ask-first setting, the builder export maps Ask first to it, but DutyGraph does not confirm the runtime applies it. The runtime control table on the issue screen names a "DutyGraph gateway approval gate"; that gate is not built. See Agent gateway.
For auditors
What does the auditor get?
A register of every agent requested through DutyGraph, including drafts never requested. In the sample, identities "found outside DutyGraph" are illustrative data; reconciling with your real enterprise inventory is Planned. A trace from each agent to its owner, Task Card, rules, decisions and gateway calls. Verify chain, which recomputes the ledger hashes. CSV and JSON exports with the ledger slice. All Built. A full zip packet with a standalone verifier script is Planned.
Does the activity log cover everything an agent does?
No. It covers calls made through the DutyGraph gateway. Calls an agent makes some other way are not in it. Run tracking from OpenTelemetry traces is Planned, and it will be evidence only.
For employees
Do employees need an account?
Not to answer questions. A team ask or an agent plan opens from a private link, with no account. To request an agent or to approve one, a person signs in to DutyGraph. Signing in through the company's own identity provider is Planned.
Can I ask for an agent from my own AI chat?
Yes, in AI apps that accept a token. Four request tools are Built, with a personal token that has only the request_agents scope. OAuth sign-in, which Claude's web and mobile connectors, ChatGPT and Copilot Studio need, is Planned. See Request from your AI.
What works on a phone?
Every page works at phone width. Answer pages are built for phones first, with voice or typing. On a phone, an approver gets one bar with the three decisions. Advisors can get phone notifications when the server has push keys set. A long access export is easier to load on a computer. See Phones and notifications.
Data and AI
Where does my data live?
The hosted service runs on Vercel, with its database on PostgreSQL (Neon) in a US East region. The database isolates each workspace's data with forced row-level security. Holdings never become Repo files. Company AI keys are stored encrypted. Agree region, retention and deletion before you load real data. See Security and data handling.
Does it use AI, and where?
Yes, through the Claude or OpenAI keys the advisor saves. Most features use the key approved for that company; some older OpenAI features use the firm's workspace key. See How DutyGraph uses AI. AI extracts research findings, polishes answers on "Here's what we heard", drafts Task Cards, reads kickoff transcripts and runs the DutyGraph Agent. It only proposes; a person accepts. No AI call is on the governance path: the evaluator, approval chain, ledger and gateway are code. Opportunities, the agent builder and the agent plan are code too.
What is planned next?
OAuth sign-in for AI connectors comes first. Then playbooks, hosted agents on the company's own vendor account, run tools and run tracking. See Roadmap and Where DutyGraph is going.
INFO
Status. Governance, the MCP request tools, discovery, Opportunities and the agent builder are Built. Identity systems, business systems and the agent runtime are Simulated in the samples. Task-based justification of each scope is Partial. Live identity connections, SSO, OAuth for AI connectors, provisioning, Decision Export, owner-departure suspension, run tools and run tracking are Planned.
Source: gitbook/product/reference/faq.md in the DutyGraph repository.