Skip to content

Panels & Orchestrators

When one agent isn’t enough, the Studio offers two coordination nodes that convene a roster of agents. Both read their members from Membership edges pointing at Agent reference nodes — you don’t list members in config, you wire them on the canvas.

A Panel runs a peer-to-peer deliberation. Inspired by MAI-DxO-style diagnostic panels, every member sees a shared transcript and contributes in turn.

  1. For each round (up to max_rounds, default 3), each member speaks in roster order, seeing the shared transcript so far.
  2. After each round, the moderator LLM reads the transcript and decides whether to stop.
  3. The moderator synthesizes the final conclusion.

Config: model (moderator), max_rounds, moderator_prompt.

Use it for: diagnosis and deliberation where multiple perspectives should see each other and build on the discussion — e.g. a hospital-simulation panel of specialists.

An Orchestrator runs a planner-led hierarchy. Members never see each other; the planner is the only one with the full picture.

  1. The planner LLM receives the task and decomposes it.
  2. It delegates sub-tasks to member agents (up to max_delegations, default 10).
  3. If adaptive is on, it re-plans based on results before delegating again.
  4. It synthesizes the final answer from the collected results.

Config: model (planner), max_delegations, planning_prompt, adaptive.

Use it for: task decomposition where one coordinator routes work to isolated experts — e.g. “research this, then draft, then fact-check” handed to specialized agents.

AspectPanelOrchestrator
TopologyPeer-to-peerHierarchical
VisibilityShared transcript, all see allMembers isolated; only planner sees all
Coordinator roleModerates, decides stop, synthesizesPlans, delegates, re-plans, synthesizes
FlowRound-robin speechesDelegated sub-tasks
Best forDeliberation / debateDecomposition / routing

Implementations live in backend/src/agents/panel.rs and backend/src/agents/orchestrator.rs.