---
url: /handbook/start-here/what-is-dutygraph.md
description: >-
  DutyGraph in plain English: an evidence-backed model of how a company works,
  and the delegated authority of the AI agents built on that work.
---

# What is DutyGraph

DutyGraph builds a living, evidence-backed model of how a company actually works, readable by people and by AI. It uses that model to find where work can improve and to decide what authority an AI agent should have.

## The thesis

DutyGraph connects one chain:

> **Business intent → Actual work → Human responsibility → Required capabilities → Technical authority → AI agents → Observed outcomes**

Agents need more than data access. They need a confirmed description of legitimate work, the systems that work uses, its rules and exceptions, an accountable owner, explicit and limited authority, human review where needed, traceable decisions, and revalidation when the company changes.

Not every link in the chain is built. The links from business intent to human responsibility are **Built** for a top-down engagement. The link from work to required capabilities is **Partial** today. Observed outcomes are **Partial**. See [Built and planned](built-and-planned.md).

## The whole product in one picture

```mermaid
flowchart TD
    subgraph OI["Operating Intelligence: how the company works"]
      R["Company research"] --> SP["Sponsor packet"] --> K["Executive kickoff"] --> TA["Team asks"]
      TA --> M["Company Repo, Task Cards<br/>and Work Map"]
      M --> O["Opportunities and Route"]
    end
    subgraph DAG["Delegated Authority Governance: what an agent may do"]
      B["Agent builder<br/>a manifest for one owner"] --> J{"Evaluator<br/>Task Cards, holdings,<br/>System Catalog, rule packs"}
      J --> AC["Approval chain<br/>manager, role owner,<br/>Internal Controls, IAM"]
      AC --> L["Ledger<br/>hash-linked record"]
      AC --> IAM["Customer's IAM<br/>provisions the agent<br/>outside DutyGraph"]
      AC --> G["Optional gateway<br/>checks routed calls"]
    end
    O --> B
    L --> AE["Audit evidence"]
    G --> AE
    AE -.->|"revalidation, mostly Planned"| M
```

## Two halves that need each other

| Half | What it covers | Built today |
| --- | --- | --- |
| **Operating Intelligence** | Research, strategy, executive discovery, employee interviews, the org map, duties and [Task Cards](../discovery/task-cards.md), workflows and handoffs, rules, constraints, opportunities and outcomes | Top-down discovery, the Company Repo, Task Cards, the Work Map, Opportunities, Route and Outcomes |
| **Delegated Authority Governance** | [Holdings](../governance/holdings.md), the business capabilities a task needs, entitlement mappings, policies, separation of duties (SoD), manifests, approval chains, audit evidence, provisioning handoffs, runtime control and revalidation | The evaluator, SoD checks, the approval chain, the ledger, the optional gateway and the evidence export. Capability mappings, provisioning handoffs and broad revalidation are Planned. |

The halves feed each other. Discovery gives governance the business context behind a permission. Governance gives discovery a practical reason to stay current: an agent's authority should be re-checked when the work, the owner or the rules change.

That feedback loop is the design, not something that runs on its own today. It has to be built through dependency tracking and revalidation. Today a saved agent manifest turns **Out of date** or **Broken** when the Repo files it cites change, and an issued agent stops at its expiry. Re-checking issued agents when holdings or Task Cards change is **Planned**.

## Three ways in

A company can start from its strategy (top-down), from one department (pod-down) or from one agent request (bottom-up). All three fill the same model. Today's built path is top-down. Read [Three ways in, one model](three-ways-in.md).

## How a top-down engagement runs

1. **[Company research](../discovery/company-research.md).** DutyGraph reads public sources. A finding stays **Inferred** until a person at the company confirms it.
2. **[Sponsor packet](../discovery/sponsor-packet.md).** The client sponsor corrects the company picture, uploads the team list and chooses who joins the kickoff.
3. **[Executive kickoff](../discovery/executive-kickoff.md).** DutyGraph builds the agenda and invitations. The advisor pastes the transcript in.
4. **[Team asks](../discovery/team-asks.md).** Each person answers questions about their own work on a phone.
5. **[Company Repo](../discovery/company-repo.md), Task Cards and [Work Map](../discovery/work-map.md).** The answers become records with an evidence state on each.
6. **[Opportunities](../opportunities/opportunities.md) and [Route](../opportunities/route.md).** Code shows hours per Task Card, proposes agents and suggests the cheapest change that could work.
7. **[Agent builder](../agents/agent-builder.md).** The advisor opens a proposed agent. The builder fills each field from the Repo and names its source.
8. **[The key join](../governance/the-key-join.md) and the [approval chain](../governance/approval-chain.md).** The evaluator proposes a state for each scope. The manager, the role owner, Internal Controls and IAM decide in turn. Each decision goes into the [ledger](../governance/ledger.md).
9. **[Audit evidence](../governance/audit-evidence.md).** An auditor can trace an agent to its owner, Task Cards, rules and decisions.

Nothing becomes a fact until a person accepts it. Code proposes; a named person decides.

## What the Task Card is for

The **Task Card** is the join between the two halves.

| In mapping | In governance |
| --- | --- |
| The record of what a person does, step by step | The intended record of why an agent needs each permission |
| Captured from the person's own words, then confirmed | Pinned by record version and content hash when access is requested |
| Marks each step human, AI or shared | Says which steps the agent may do and where it stops for a person |
| Names the systems the work touches | Feeds the System Catalog |

::: warning
**The target, not yet the fact.** The aim is that a confirmed Task Card justifies each permission an agent holds. Today that is **Partial**. The evaluator keeps an agent within its owner's holdings and the company's rules. It does not yet limit the requested scopes to what the selected Task Cards need. See [The key join](../governance/the-key-join.md).
:::

## How DutyGraph fits with your identity systems

DutyGraph complements a company's existing identity and access management (IAM). DutyGraph records the business justification and the authorization decision. The company's own IAM and provisioning tools create the agent's identity and grant its access.

* It does not connect to identity systems today. It reads an uploaded access export.
* It does not provision the agent identity. It gives IAM a "For IAM to do" card.
* It does not run agents. It exports files formatted for the company's own AI tools. See [Exports and runtimes](../agents/exports-and-runtimes.md).
* Its gateway is optional and narrow. It checks only the calls routed through it.

## Who runs it today

DutyGraph is advisor-led today. An advisor or a forward-deployed engineer (FDE) operates it for each company. Client people take part through private links, and most need no account. An employee can also request an agent from their own AI chat over the Model Context Protocol (MCP). See [Who uses DutyGraph](who-uses-dutygraph.md).

::: info
**Status.** Top-down discovery, the Company Repo, Task Cards, the Work Map, Opportunities, Route, the agent builder, the evaluator's ceiling and rule checks, SoD checks, the approval chain, the ledger, the gateway check and the evidence export are **Built**. Limiting an agent's scopes to what its Task Cards need is **Partial**. The identity systems and the business systems behind the gateway are **Simulated** in the samples. Pod-down and bottom-up entry, step-level capability mappings, live identity connections, provisioning, broad revalidation and running agents are **Planned**.
:::
