The short answer

Persistent human-agent collaboration keeps tasks, outputs and human decisions in durable shared state so work can continue across turns and process restarts. In AgentPlat, an Agent Room organizes that work around a shared objective.

What needs to persist beyond a conversation?

A useful agent workflow often lasts longer than a single response. Research feeds a draft, a person requests a revision, and another task depends on the approved result. A transcript can describe these events, but the application also needs to know which output was reviewed and what work remains.

Persistent collaboration makes those relationships explicit. Store task identity, dependencies, artifact versions and review decisions alongside the conversation. A returning participant should be able to distinguish a proposal from an accepted deliverable without reconstructing the entire chat.

How Agent Rooms organize shared work

AgentPlat is an open-source TypeScript framework that adopters run on their own infrastructure. Its Agent Rooms provide a shared work container for human and agent participation. The persistent collaboration guide demonstrates research, drafting, a revision request and approval through the Rooms HTTP API.

In that example, the retained record includes completed tasks, research and proposal artifacts, proposal versions, approval records and event history. These are inspectable application records, rather than details that only exist in a model's conversational memory.

An example: research, draft, review, revise

  1. Define the objective. Create a Room for a proposal with a clear deliverable and review expectations.
  2. Gather evidence. A research task produces an artifact that the drafting task can use.
  3. Produce a version. The drafting task creates a proposal linked to the work that informed it.
  4. Request a revision. Record the review decision and preserve the earlier proposal.
  5. Review the successor. Approve the specific version that satisfies the objective.

The documented demo scripts the human API role. It demonstrates the retained workflow; it is not a ready-made review interface or a study of live human collaboration. Your application supplies the user experience and connects authenticated participants to their allowed actions.

When should you use persistent collaboration?

Use this model when work needs repeated review, multiple contributors or a durable handoff between sessions. Proposal preparation, internal research and document revision are examples where artifact identity and review history matter.

A short request with one disposable result may need much less machinery. A predefined sequence may fit a step-based workflow. The adoption guide explains how to choose a work model. Start with the smallest representation that preserves the decisions your application must explain.

Persistence, approval and recovery are separate concerns

Saving a result does not by itself authorize an external action. Reviewing a proposal and permitting a provider to change another system require distinct controls. Likewise, a durable artifact does not prove that every interrupted provider request can be resumed safely.

Test the boundaries separately: inspect a revised artifact, attempt an action without approval, and stop a worker at a documented recovery point. Keep the observed result attached to the scenario you actually ran. AgentPlat's examples provide concrete starting points, while production behavior depends on your provider integration, storage and operational controls.

Sources and further reading

Documentation reviewed . Consult the linked documentation for current implementation details.