Skip to content

The DutyGraph loop ​

DutyGraph is designed around one loop. It learns how the company works, finds what to improve, authorizes the change, watches what happens and updates its model. Then it starts again. Some stages are built today. Others are Planned.

WARNING

The loop describes where DutyGraph is going, not only what it does today. Read the Status column before you describe a stage to a client. Do not present a Planned stage as available.

The loop in one picture ​

mermaid
flowchart LR
    S1["1. Discovery<br/>and mapping"] --> S2["2. Improvement<br/>opportunities"]
    S2 --> S3["3. Preparation and<br/>authorization"]
    S3 --> S4["4. Execution through<br/>approved systems"]
    S4 --> S5["5. Observation and<br/>measurement"]
    S5 --> S6["6. Updating the<br/>operating model"]
    S6 -.->|repeat| S1
    S5 -.->|"monitoring and revalidation<br/>continue during and after execution"| S3
    classDef built fill:#d9f2e3,stroke:#2e7d4f,color:#123;
    classDef partial fill:#fff3cd,stroke:#a07800,color:#321;
    classDef planned fill:#eeeeee,stroke:#888,color:#333,stroke-dasharray: 4 3;
    class S1,S2 built;
    class S3,S5,S6 partial;
    class S4 planned;

Green stages are Built. Yellow stages are Partial: some of the stage is Built, the rest is Planned. Grey stages are Planned.

Authorization comes before governed execution. A person approves an agent's authority before the agent acts. Governance does not stop there. Checks, measurement and revalidation continue while the agent runs and after it stops.

Stage by stage ​

StageWhat it meansWhat DutyGraph does todayStatusWhere in the app
1. Discovery and mappingLearn how the company works and record it: people, duties, Task Cards, systems and handoffs.Company research, the sponsor packet, the executive kickoff (transcript is paste only) and team asks on a phone. The Company Repo holds the result with an evidence state on each file. The Work Map shows how work moves. This runs top-down today. See Three ways in, one model.BuiltOverview, Discovery, Repo, Work Map
2. Improvement opportunitiesFind waste, bottlenecks and work that suits an agent. Choose the right change, which is often not software.Opportunities shows hours a week per Task Card, proposed agents and possible constraints. Route ranks a ladder of 9 changes, from "No change" to "Hiring or outsourcing". Code makes the proposals. It shows hours only, with no money and no return on investment (ROI).BuiltOpportunities, Opportunities › Changes, Strategy
3. Preparation and authorizationPrepare the change and decide what authority an agent gets, before it acts.The agent builder writes a versioned manifest. The evaluator checks each scope against the owner's holdings, the System Catalog and the rule packs, and runs SoD checks. The approval chain decides, and the ledger records each decision. Limiting scopes to what the Task Cards need is not enforced yet. The company's IAM provisions the agent outside DutyGraph.PartialAgents › Workspace, Requests, Review queue, Rules
4. Execution through approved systemsPeople, software, agents, vendors or engineers do the work, within the authority approved in stage 3.DutyGraph runs no change and no agent. It exports agent files formatted for Claude Code, Claude Managed Agents and OpenAI Agents, and a draft work order. An optional gateway checks calls routed through it; the business systems behind it are Simulated.PlannedAgents › Saved agents, Issued agents
5. Observation and measurementWatch what happened and compare it with what was predicted and approved.The gateway records each routed call. Outcomes shows each change's prediction, measures, readings and review. A result reads "Supported by 1 pilot · not proven", never "proven". Run tracking of what an agent actually did is Planned.PartialOutcomes, Agents › Audit, Reports
6. Updating the operating modelFeed what happened back into the model, and re-check authority that depends on what changed.Accepted answers, decisions and outcome reviews become new records, and the Repo rebuilds from them. A saved manifest turns Out of date or Broken when the files it cites change. A saved report shows "Out of date" when the Repo changes. Re-checking issued agents when holdings or Task Cards change is Planned.PartialRepo › To review, Reports

How these stages map to the vision's ten steps ​

The long-term vision names ten steps. These docs group them into the six stages above, and place Govern before Execute.

Vision stepStage in these docs
Discover, Map1. Discovery and mapping
Diagnose, Route2. Improvement opportunities
Govern3. Preparation and authorization, and continuing checks in stages 4 to 6
Execute4. Execution through approved systems
Measure5. Observation and measurement
Learn, Update6. Updating the operating model
OperateAcross every stage: DutyGraph becoming part of normal work. Planned.

How the app sections follow the loop ​

The company sidebar reads: Overview, Discovery, Repo, Work Map, Opportunities, Agents, Outcomes, Strategy, Reports.

  • Discovery gathers the answers. Repo holds what was learned. Work Map shows how the work runs today.
  • Opportunities says what to change. Route lives here, as the Changes tab.
  • Agents builds agents and governs their access.
  • Outcomes shows whether a change worked. Strategy looks ahead.
  • Reports groups its report catalog by loop step: Discover, Map, Diagnose, Route and Measure.

Direction: the model improves because the business operates ​

WARNING

This section is direction, not a description of the product. Nothing in it is built.

Discovery is meant to be the start, not the end. The intended learning loop during execution is:

Exception → escalation → a person resolves it → evidence captured → proposed rule or Task Card update → owner review → evaluation → approved new version → revalidation of the agents that depend on it.

The aim is that each meaningful escalation can improve the next case. The model would then get more accurate because the business operates, not only because an advisor runs another round of discovery. See Where DutyGraph is going.

INFO

Status. Discovery and mapping (top-down) and improvement opportunities are Built. Preparation and authorization is Partial: the evaluator, SoD checks, approval chain and ledger are Built; limiting scopes to Task Card needs and provisioning are Planned. Observation and updating are Partial. Execution, run tracking, revalidation of issued agents and the learning loop are Planned. Identity and business systems in the samples are Simulated.

Source: gitbook/product/start-here/the-dutygraph-loop.md in the DutyGraph repository.

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