---
url: /handbook/roadmap/vision.md
description: The long-term direction for DutyGraph, and which parts of it exist today.
---

# Where DutyGraph is going

This page describes where DutyGraph is headed: a living, evidence-backed, AI-readable operating model of a company. For each part of that direction, it says what you can use today and what is still vision.

::: warning
**This page is direction, not shipped capability.** Anything this page calls vision is not built. Where a capability exists today, the page links to the page that describes it. See [Built and planned](../start-here/built-and-planned.md).
:::

## The one-sentence vision

> DutyGraph becomes the living, AI-readable operating model of a company. It learns how work actually happens, captures the judgment that governs that work, prepares work for people, software and AI agents, decides with people what authority each agent should have, and learns from the outcomes.

## The core thesis

Models keep improving, and agents are becoming common. The scarce input is company-specific operating knowledge: who owns the work, which rules apply, which exceptions are legitimate, and who decides when the rules are not enough.

> **Agents do not only need access to company data. They need approved operating judgment and evidence-backed authority.**

DutyGraph connects one chain: **business intent → actual work → human responsibility → required capabilities → technical authority → AI agents → observed outcomes.**

Two principles follow:

* **AI proposes. Code checks. Accountable people approve.** No model silently owns the company's truth.
* **Understanding is not permission.** That an AI can describe a process does not mean it may act in it.

## Operating intelligence is the shared platform

DutyGraph is not only an identity product. Its foundation is **operating intelligence**: the company's goals, duties, Task Cards, workflows, handoffs, rules, constraints, opportunities and outcomes, each with its evidence.

Two halves draw on it:

* **Operating Intelligence** helps leaders see how the company works and what to change.
* **Delegated Authority Governance** decides what authority an agent should have, with evidence for each decision.

Discovery gives governance the business reason behind a permission. Governance gives discovery a reason to stay current. That feedback has to be built through revalidation; it does not run on its own today.

The model must keep apart what was **observed**, **inferred**, **confirmed** and **approved**. A confirmed observation is not authority. A hash-linked record shows the record was not changed; it does not prove the business claim inside it.

**Today:** the [Company Repo](../discovery/company-repo.md) is the built start of this model, with [Task Cards](../discovery/task-cards.md) as its core unit of work.

## Evidence-backed delegated authority

**Vision:** each permission an agent holds is traced from a confirmed task step, to the business capability it needs, to a technical entitlement an application owner validated, to an approved grant. The company's own IAM tools provision it. DutyGraph records the justification and the decision.

**Today:** the evaluator keeps an agent within its owner's holdings and the rules; the approval chain, ledger and optional gateway are Built. Limiting scopes to what the Task Cards need is **Partial**. Step-level capabilities, validated mappings, a Decision Export to the company's IAM and an Event Import back are **Planned**. See [The key join](../governance/the-key-join.md).

## Many ways in, one model

**Vision:** a company can start from its strategy, from one department or from one agent request. All three fill the same model through one engine.

**Today:** top-down discovery is Built. Pod-down discovery and bottom-up task capture are **Planned**. See [Three ways in, one model](../start-here/three-ways-in.md).

## The loop

The vision is one loop, not a set of separate products: **discovery and mapping → improvement opportunities → preparation and authorization → execution through approved systems → observation and measurement → updating the model**. Authorization comes before governed execution. Monitoring and revalidation continue during and after it. See [The DutyGraph loop](../start-here/the-dutygraph-loop.md).

## Continuous revalidation

"Living" should mean that a change in authoritative information re-checks the authority that depends on it, not only that files can be refreshed.

| Change | Intended response | Today |
| --- | --- | --- |
| The owner's holdings change | Find affected agents, recompute, and narrow or suspend unsafe authority | Planned |
| A Task Card changes materially | Agents that rely on the old version are flagged for reassessment | Partial: saved manifests turn **Out of date** or **Broken** |
| An authorization expires | Calls stop at every supported check point, and the stop is recorded | Built at the gateway |
| The owner leaves or changes role; rules, mappings, implementations or applications change; an unmanaged agent is found | Reassess the affected agents | Planned |

## Portable projections of the Repo

The Company Repo is compiled from validated records. **Vision:** it can be compiled into views for different readers, each under explicit privacy and permission rules:

* an **executive** view: goals, key duties, risks, contradictions and opportunities;
* an **agent context** view: task context, rules, handoffs, instructions and approved tools;
* a **governance evidence** view: task-to-scope justification, rule and approval references, and lifecycle; and
* a **portable company export**.

Sensitive holdings would never be printed into a downloadable tree. **Today:** **Download your company** exports the Repo. The other views are Planned.

## From task to agent

**Vision:** DutyGraph compiles an approved task into what a builder needs: instructions, task context, input and output formats, approved tools, constraints, human checkpoints, test cases and version hashes. It builds on today's exports, not a second manifest format.

Four things stay separate:

| Thing | Question |
| --- | --- |
| **Agent Definition** | What should the agent do? |
| **Agent Implementation** | Which instructions, tools and code does it use? |
| **Agent Authorization** | What authority does it have? |
| **Agent Execution** | What did it actually do? |

A draft grants no authority. A material change to the implementation is revalidated against its authorization.

**Today:** the [Agent builder and manifests](../agents/agent-builder.md) and [Exports and runtimes](../agents/exports-and-runtimes.md). DutyGraph does not run agents. Agent Build Packets and Agent Studio are vision.

## Operate: part of normal work

**Vision:** DutyGraph becomes part of normal work. **Meeting OS** replaces most status meetings with a short async **DutyGraph Pulse**, and every meeting updates the model.

**Today:** answer pages built for phones, advisor notifications, scheduled research refresh and [Living DutyGraph](../advisors/living-dutygraph.md). Meeting OS and Pulses are vision.

## Learning from observed outcomes

A recommendation is not value, and a deployment is not value. DutyGraph should track each change from **Predicted** to **Observed** to **Realized**, and feed the result back into the model.

| Stage | Meaning | Today |
| --- | --- | --- |
| **Predicted** | What a change is expected to do, recorded before it starts | **Built.** [Route](../opportunities/route.md) records a prediction. Predicted hours with an agent exist only in the sample companies. |
| **Observed** | What actually happened | **Built** for measures in [Outcomes](../opportunities/outcomes-and-strategy.md). **Planned:** run tracking of what an agent actually did. Observed data is evidence. It never grants or blocks access. |
| **Realized** | The lasting effect, fed back into the model | **Vision.** |

**Vision:** escalations and resolved exceptions become evidence. DutyGraph proposes a rule or Task Card update, the owner approves a new version, and the agents that depend on it are revalidated.

## How we judge a new feature

Does it make DutyGraph a more accurate, useful, executable, governed and self-improving representation of how the company operates? If not, it is probably a distraction.

::: info
**Status.** This page is direction. Top-down discovery, the Company Repo, the evaluator's ceiling and rule checks, the approval chain, the ledger, the gateway and outcome measures are **Built**. Limiting scopes to Task Card needs and revalidation are **Partial**. The identity and business systems in the samples are **Simulated**. Pod-down and bottom-up entry, step-level capabilities, Decision Export, Event Import and Repo projections are **Planned**. Meeting OS, DutyGraph Pulse, Agent Build Packets, Agent Studio and learning from escalations are vision.
:::
