Appearance
Exports and runtimes
DutyGraph writes an agent out as files for the AI tool a company already uses. Each file is "formatted for" a runtime. DutyGraph does not load the file, run the agent or host it.
There are two exports:
- The builder export. Any saved manifest, from Agents › Workspace or Agents › Saved agents.
- The governed runtime export. An issued agent, from the issue screen in Agents › Review queue.
The builder export
Select Export manifest (or press E), or Download .zip in the builder. Choose one format.
| Format | Files | Formatted for |
|---|---|---|
| Manifest sheet | manifest.md. Print the sheet gives a light page to print or save as PDF. | People to read and print |
| DutyGraph manifest JSON | manifest.json | dutygraph.agent-manifest.v1 |
| Claude Code folder | .claude/agents/<slug>.md (the subagent), .claude/skills/<slug>/SKILL.md (the skill), settings.fragment.json | Claude Code |
| Claude Managed Agents JSON | managed-agent.json (the agent definition), skills/<slug>/SKILL.md | A Claude Managed Agents agent definition |
| OpenAI Agents JSON | openai-agent.json | An OpenAI Agents API agent definition (beta, agents=v1) |
Every format also has README.md and GAPS.md. The download is the saved version, as <slug>-v<version>-<format>.zip. With unsaved changes, the button reads Save and download .zip.
README.md and GAPS.md
- README.md says how to use the files, who provides credentials, and where to send each
[S]and[H]step. - GAPS.md lists every value that is Not found, with the person to ask and one question. DutyGraph never fills a gap. It leaves the value out of the runtime files. An export with gaps is labelled "Draft · N gaps".
- Every file carries the manifest id, version,
manifestHashand the Repo fingerprint.
What each runtime file allows
DutyGraph maps each tool policy to the runtime's own setting. Where the runtime has no matching setting, the export picks the tighter one.
| Runtime | How the policies map |
|---|---|
| Claude Code | Allow, Ask first and Off become allow, ask and deny rules. "Auto within set limits" becomes ask. The subagent can read, search and list files. It cannot run commands, change files or use the web. The settings fragment denies reading .env files, key files and a secrets folder. |
| Claude Managed Agents | Allow stays allow. Ask first and "Auto within set limits" become always ask. Built-in tools are limited to read, glob and grep. You must choose a Claude model in Agent config first. |
| OpenAI Agents | The file lists only tools DutyGraph allows, because this API has no ask-first setting. The README names every tool left out. DutyGraph does not choose a model, so you set model. |
A tool reaches a runtime file only when the connector, its server and its tool names are known. Until then, the tool is left out and listed in GAPS.md.
When DutyGraph refuses a builder export
| Case | Result |
|---|---|
| The manifest is Broken | Only the sheet and the JSON export. Review and save it again first. |
| No model chosen | No Claude Managed Agents file. Choose a Claude model. |
| A file holds a value that looks like a secret | Refused. The message names the file. |
| A file holds an email address or phone number | Refused. Remove it from the Task Card. |
| A file holds interview or answer words | Refused. |
The governed runtime export
After the approval chain decides, IAM opens the issue screen. Runtime export offers Claude, ChatGPT and Copilot. For each, a fixed table shows each control, whether it is mandatory, and whether DutyGraph marks that runtime as able to apply it. The table is a recorded assessment. DutyGraph does not test the runtime.
| Mandatory control | What actually applies it today |
|---|---|
| Only approved scopes | The DutyGraph gateway's tool allowlist, for calls routed through it |
| Ask before supervised calls | Only a runtime's own approval setting, where it has one. The table's note names "the DutyGraph gateway approval gate". That gate is not built. |
| Log every gateway call | The gateway's activity record, for routed calls |
| Stop at the end of the agent's lifetime (rule AGT-01) | The gateway checks the end date on every routed call |
| For an event-bound agent, stop when the event ends (rule AGT-10) | The gateway checks the event window on every routed call |
If the table marks a runtime unable to apply a mandatory control, DutyGraph refuses to issue to it and refuses the export. The screen shows Export blocked and names the control. In the current build, Copilot is refused, because DutyGraph cannot show that a Copilot export stops the agent at its end date.
WARNING
The table marks "Ask before supervised calls" as supported for every runtime. Do not rely on it as a per-action approval. DutyGraph has no step where a person approves a single call. See Agent gateway.
For an allowed runtime, Export for <runtime> gives a JSON file. It carries the issue record id, the rule pins, the issued and removed scopes and the expiry. It says: "Customer IAM provisioning is not connected." The agent can reach its tools through the agent gateway, which checks only calls routed through it.
Who provisions
DutyGraph decides and records. The company's IAM provisions. The issue screen shows a For IAM to do card that names the agent identity to create, with exactly the approved scopes. DutyGraph does not create the identity in any system.
An export is an implementation file. It is not an authorization. A material change to an exported agent should be checked again against its issued authorization. That check is Planned. See Approval, provisioning and revalidation.
Planned: playbooks and hosted agents
A design for two ways to deliver the same manifest version is written down. It is not built.
| Playbook (a skill) | Hosted agent | |
|---|---|---|
| Where it runs | The employee's own AI chat, with their own access | The vendor's servers, on the company's own account |
| Limits | Guidance: a skill is instructions | Applied by the vendor's tool list, permission policy and budget, with the gateway checking calls routed through it |
| Status | Planned. The SKILL.md file in the Claude exports is Built as a file only. Serving it from DutyGraph is Planned. | Planned. DutyGraph would create the agent on the company's account only if the company opts in. The vendor runs it. DutyGraph hosts nothing. |
Also Planned: a Microsoft 365 Copilot declarative agent app package, starting runs from chat, and run tracking from OpenTelemetry data.
INFO
Status. Built: five builder formats with README.md and GAPS.md, the export refusals, the runtime control table at issue, and the governed JSON export. Not built: a DutyGraph gateway approval gate for single calls. Simulated: the agent runtime and identity systems in the governance sample. Planned: a Copilot declarative agent package, playbooks served from DutyGraph, hosted agents on the company's own vendor account, provisioning in the company's IAM, and run tracking.
Source: gitbook/product/agents/exports-and-runtimes.md in the DutyGraph repository.