Run an agent debate

Desktop can run a sequential team conversation declared in Grid. Two speakers exchange arguments in a shared room; a chair assesses the exchange and either continues or concludes it. Each speaker sees the earlier accepted public replies.

In the Process surface, choose Create process → Guided debate. Enter a process name and a quoted topic, then Apply draft in Code. Review the generated Team, Room and Process and Save. Select the saved plan, bind its three AI profiles and Start explicitly. The starter uses the same sequential conversation shown below, with up to three rounds. Creation never runs a model or saves for you. If the host does not offer starters, the source below remains available.

Open a model, switch its source editor to Author (or use Code), paste the following source, and Save. Saving prepares the process; it does not call a model.

# Desktop: save this model, then open Agent > Tasks > Processes and Start.
# Start creates a fresh room. Each speaker sees prior accepted public turns.
# Profiles and Start limits are supplied by the host; evaluating a plan does no work.

TEAM panel
  advocate USING "debate-advocate"
  challenger USING "debate-challenger"
  chair USING "debate-chair"
END

ROOM seminar USING panel

PROCESS debate(topic) IN room WITH advocate, challenger, chair
  SHARE topic
  REPEAT UP TO 3
    supporting = ASK(advocate, "Make the case, addressing the latest reply and guidance.")
    opposing = ASK(challenger, "Respond to the advocate's current argument.")
    direction = ASK(chair, "Assess the exchange and guide or conclude it.") EXPECT {
      action: "continue" | "conclude" | "unresolved",
      assessment: string,
      focus: string
    }
    YIELD direction UNTIL direction.action != "continue"
  END
END

proposal = debate("Should the library open on Sundays?") IN seminar

Open the agent panel, then Tasks → Processes. Select proposal and choose a model for each profile: debate-advocate, debate-challenger, and debate-chair. Use AI setup uses your saved provider and model; explicit selections can use different models for the three roles. Credentials and model-server addresses stay in the existing AI setup. A profile name in the source is a binding to these choices, not a separate agent-definition language.

Press Start process. The run continues when you close the panel. Select the run to inspect its status and ordered conversation as replies finish. Room metadata becomes available when the run pauses or finishes; the final result appears when it completes. Instructions reveals the prompt for a turn. The chair's action ends the loop when it is conclude or unresolved; continue begins another exchange, up to three. The result includes the last direction, the iteration count and whether the condition or the limit stopped the loop.

SHARE topic makes the topic available to the participants. The room supplies the accepted conversation automatically. The statements run in source order: advocate, challenger, chair. EXPECT checks the chair's response before accepting it; an invalid response stops the run instead of supplying an unchecked decision.

Use Cancel process to stop active work. Already accepted replies remain visible. An in-flight call may still finish; a late reply is marked unaccepted and cannot advance the discussion. If a Start response is uncertain, Retry same start reconciles that request without creating another run. After a successful Start, pressing Start process again deliberately creates a new run. Saving changed source does not rewrite an already running process.

Current boundaries

This is the sequential Desktop subset on the 0.66.8 development branch. Other product packages do not enable it by default. Model calls use the configured account and its normal costs. The local product walkthrough used a deterministic model server to establish orchestration, not the quality of real model judgments.

Replies appear after each turn finishes, while later speakers are still working. The panel follows the latest turns; use Previous turns to read earlier pages and Follow latest turns to resume following. This is completed-turn display, not token streaming. History is bounded: if it exceeds retention, the panel states that it was withheld rather than presenting a truncated conversation as complete. On macOS and Linux, Desktop retains runs in a private recovery store: accepted turns survive a host restart, and a restored run waits for an explicit Resume. Windows Desktop and hosts without a selected store keep runs volatile across restart. Closing the panel does not stop a run.

This example does not require external tools. API reads have a separate walkthrough. Browser operations, long-lived room handles, dynamic swarms and distributed agents remain unfinished. Those design samples are not promises of executable product support.