Agent Communication Topology
In a multi-agent system, who can talk to whom is a design axis distinct from how many agents you run or what tier they are. The source draws the two canonical points:
- Subagents — hub-and-spoke (star). One main session delegates to worker subagents; each worker serves the main session and reports back to it, but the workers cannot talk to each other. The orchestrator is the only hub. Cheaper, and the default when you spin up subagents.
- Agent teams / councils — mesh. Agents can talk to the main session and to each other, and can debate — "argue until they reach some sort of consensus." More expensive, because inter-agent chatter is more tokens and more turns.
The distinction has a practical consequence: cross-perspective critique can happen between the agents (a mesh debate) or be orchestrated by the hub as a separate stage over the workers' independent outputs. STORM takes the second route — five lens subagents that never talk to each other, followed by an explicit hub-run contradiction map + peer review, which delivers the blind-spots-caught payoff of a council without paying for a mesh. So topology is not destiny: a star plus a review stage can substitute for a mesh debate.
This is orthogonal to Bounded Fan-Out (the count axis) and Model-Tier Routing (the tier axis). It is also distinct from Agent Task Graph, which is the coordination substrate a mesh team needs (shared task list, dependency edges, a channel to communicate) rather than the yes/no of whether workers may communicate at all.
Claims
- The communication topology of a multi-agent system — workers that talk only to the orchestrator (star) vs agents that also talk to each other (mesh) — is a design axis separate from agent count and model tier. principle — durable: it determines where cross-perspective interaction can occur and is chosen independently of how big or how expensive the roster is.
- Subagents are hub-and-spoke: they serve one main session and report back to it, but cannot communicate with each other. observation — the source's framing of the subagent model; the main session is the sole hub and scheduler.
- Agent teams / councils are mesh: agents can talk to each other as well as the main session, and can debate to consensus — at materially higher cost. observation — "agent teams are much more expensive than subagents"; the extra cost is the inter-agent conversation.
- Reach for a mesh team only when you need agents to cross-examine each other's reasoning; use star subagents for independent parallel work the hub aggregates. best practice — context: choosing topology by task. Debate-to-consensus earns its cost on genuinely contested decisions; for decomposable research a star is enough. See Bounded Fan-Out, Agentic Simplicity.
- A star topology can still get cross-perspective critique by adding an explicit hub-run cross-review stage after the parallel workers finish — you don't need a mesh to get the "blind spots the council caught" payoff. principle — durable: Adversarial Planning Council's value comes from the peer-review round, and that round can be a hub stage over star-topology outputs rather than a mesh argument. Cheaper, and it keeps the review a controlled step.
Related
- Adversarial Planning Council — a council can be realized either as a mesh debate or as star-topology parallel personas plus a hub-run cross-review; the payoff (blind spots surfaced) is in the review round either way.
- Agent Task Graph — the coordination substrate a mesh team needs (shared task list, dependency edges, a communication channel); the star needs only ping-back to the orchestrator.
- Bounded Fan-Out — the orthogonal count axis: topology says whether they talk, fan-out says how many there are.
- Model-Tier Routing — the orthogonal tier axis: which model each agent carries, decided independently of the topology.
- Workflows vs Agents — a star of workers on predefined roles is workflow-shaped; a mesh that debates its own way to consensus leans agent-shaped.
- Layered Agentic Architecture — subagents/teams are the "agent = scale" layer; topology is how that layer is wired internally.
- Agent Supervision — in both topologies the human still supervises the orchestrator; a mesh adds agent-to-agent influence the human doesn't directly see.
- Distillate: STORM: A Fixed Panel of Adversarial Research Lenses, Packaged as a Skill