Skip to content

Roadmap ​

This section lists the work that is written down but not built, and links to the long-term direction.

WARNING

Everything in this section is Planned or vision. It is written down, not built. Dates are not committed. For what exists today, see Built and planned.

ItemWhy it matters
Task Card requirement check (rule CEIL-02) and step-level capability evidenceThe correctness gap. Today the evaluator keeps an agent within its owner's holdings and the rules, but does not limit scopes to what the Task Cards need. This adds that limit, with business capabilities per step and entitlement mappings validated by application owners. Missing evidence fails closed.
Suspend agents when their owner leavesToday marking a person as left does not suspend their issued agents or credentials. This suspends or revokes them, records it in the ledger, and proves calls are blocked afterwards.
Per-action approval for supervised callsToday the gateway has no per-action human approval. This adds a gateway approval gate, or relies only on a runtime's own approval feature where one is confirmed to exist.
Decision Export and Event ImportSend the approved decision, with its justification and evidence, to the company's IAM or IT service management tools. Bring back events such as provisioning done, access changed, credentials revoked or owner left, and reconcile them with what was approved.
Agent inventory reconciliationMatch the company's real agent identities to the agents governed in DutyGraph. Flag agents outside DutyGraph, approved agents missing outside, and differences in access.
Dependency revalidationRe-check issued agents when holdings, Task Cards, rules or mappings change. Today only manifest freshness and expiry are built.
Pod-down discoveryLet a department head start discovery for one unit without a full company map. See Three ways in, one model.
Agent Authority RecordOne view of an agent's purpose, owner, Task Cards, scope calculations, SoD findings, decisions, issued authority and lifecycle, computed from existing records.
Just-in-time task captureGovern an agent before the whole company is mapped. Capture only the Task Card the request needs.
Employee Knowledge CompensationPay employees for the knowledge in their accepted Task Cards and for measured improvements, with KAR (Knowledge Above Replacement), agent ownership, protection tiers and secret handling. Per-card pay rules exist in Earn today.
Workflow GainsharingShare the measured gain with the employee who runs a Task Card faster or produces more: run records that end at the hand-off, work time and wait time per link, a coach, a frozen baseline and a choice between time back and more output. A time-saved bonus rate exists in Earn today.
The workflow coach and the AI use checkAn AI coach in the employee's app with four aims: faster, more, cheaper, safer. It asks which AI a person uses for each step, checks cost against a live price list, tests a cheaper setting on the real task, and compares data use with the company's AI policy. Also: open requests with wait time, and notifications across push, email and text. The role AI persona it would use is built.
Always-on DutyGraph AgentOptional. Business Autopilot runs on its own: it wakes on a schedule or a change, checks the map, remembers what it asked, and drafts a few questions for the advisor to approve. Today the advisor clicks Run a pass.
Key-person risk mapOptional add-on after mapping. A page that scores who the business depends on from the Task Cards, with a counts-only link an owner can forward to a buyer or banker. A prototype was built and not merged. The Work Map's Key-person risk lens is built.
Private deployment and private AIOptional add-on service. Four tiers: hosted on the client's own Azure, AWS or Google AI contract; installed in their cloud account; on a model they host; fully offline. The Azure OpenAI, Bedrock and Gemini connections are built but not tested on a live account.
OAuth sign-in on the MCP serverLets web and mobile AI connectors add DutyGraph. Token sign-in works today.
Live identity and IGA connectionsRead holdings, rulesets and certifications from Okta, Entra ID and Saviynt, read-only, set once per company.
SSO for reviewers and employeesReviewers decide under the company's own sign-in.
Provisioning handoffCreate the access request in the company's IT service management tool instead of a "For IAM to do" card.
Run tools and run trackingStart an approved agent from chat. Record what it actually did next to what it was approved to do, from OpenTelemetry data. Observed data is evidence. It never grants or blocks.
Hosted agents on the company's own vendor accountDutyGraph creates the approved agent on the company's account. The vendor runs it. The gateway checks every call. DutyGraph hosts nothing.
Standalone evidence verifierRe-check an export's ledger slice without DutyGraph.

Now, Next, Later ​

This view shows order, not dates. It comes from the employee-agent plan, whose phases are named E1 to E7, and from the current work list. Each item is Planned.

Now ​

ItemWhat it adds
Task Card requirement check (CEIL-02) and step-level capability evidenceFirst, because it is the correctness gap. The evaluator keeps only scopes that confirmed, validated Task Card requirements support, that the owner holds, and that every rule allows.
Suspend agents when their owner leavesAffected agents are suspended or revoked in one consistent step, with ledger evidence and tests that prove later calls are blocked.
OAuth sign-in on /api/mcp (E1)An employee adds the DutyGraph connector in Claude on the web or a phone, then in ChatGPT and Copilot Studio, and signs in once. Reads return only their own work. This is the critical path for the rest.
Request tools behind OAuth (E2)The four request tools, built today with token sign-in, give the same result after OAuth sign-in.
Governance on the fictional sample companySet up governance today refuses every sample company. The fix lets the fictional sample run the governed path end to end.

Next ​

ItemWhat it adds
Playbooks (E3)An approved agent served as a playbook: instructions that run inside the employee's own chat with their own access. Each load is logged, and it is always the approved version.
Managed Agents adapter (E4)For a company that opts in, DutyGraph creates the approved agent on the company's own Claude Managed Agents account, with the gateway as its tool server. Only after the approval chain is complete. The vendor runs it.
Run tools (E5)run_agent, get_agent_run, confirm_agent_step and cancel_agent_run. Start a hosted agent from chat and get the result back. Only the person who must confirm a paused shared step may confirm it in chat. Request decisions stay in DutyGraph.
Run tracking from OpenTelemetry (E6)A receiver for agent traces. The agent's page shows runs, tool calls, stop reasons and "observed vs approved" findings. A trace that matches no known agent becomes an unmapped-agent finding.
Read-only identity connections and reviewer sign-inHoldings, rulesets and certifications read from the company's identity and governance tools instead of an uploaded file. Reviewers sign in through the company's identity provider.
Per-action approval for supervised callsA gateway approval gate, or a runtime's own confirmed approval feature, before a supervised call runs.
Agent inventory reconciliationBefore broad runtime expansion: match the company's real agent population to governed agents and record each difference as a finding.
Dependency revalidation, including recertification and driftRe-check issued agents when the owner's access, a Task Card, the rules or a mapping changes.
Decision Export and Event ImportThe approved decision goes to the company's IAM tools; their events come back and are reconciled. Approval, provisioning and observed activity stay separate states.
Agent Authority RecordOne auditable view of each agent, computed from the records DutyGraph already keeps.
Standalone evidence verifierA script that ships with an export and re-checks its ledger slice.

Later ​

ItemWhat it adds
OpenAI and Microsoft adapters (E7)The same run records and tracking for agents on OpenAI and Microsoft Copilot. Google and AWS wait for customer demand.
Provisioning handoffCreate the agent's access request in the company's IT service management tool.
Just-in-time task captureCapture one Task Card on demand when an employee requests an agent. Its place in the order is not decided.
Pod-down discoveryA department head sponsors discovery for one unit. Its place in the order is not decided.
Discovery for matrix organizationsDiscovery today is designed around a functional organization.

The long-term direction behind this list, including Meeting OS, DutyGraph Pulse, Agent Build Packets and Agent Studio, is on Where DutyGraph is going. Those items are vision, not planned work.

INFO

Status. Nothing on this page is built. The request tools, the evaluator's ceiling and rule checks, the approval chain, the gateway, the ledger and the agent exports that these items build on are Built. The identity and business systems in the samples are Simulated. See Built and planned.

Source: gitbook/product/roadmap/README.md in the DutyGraph repository.

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