Skip to content

Multi-Agent Orchestration

Commander's core differentiator is its ability to orchestrate multiple agents across 5 canonical topologies, aligned with Anthropic's "Building effective agents" ontology. Nine legacy topology names remain as aliases for backward compatibility during a 2-version migration window.

Canonical Topologies

TopologyDescriptionLegacy Alias
SINGLEOne agent handles the entire task
CHAINSequential pipeline, each agent builds on previous outputSEQUENTIAL
DISPATCHIndependent subtasks run concurrently, results synthesizedPARALLEL
ORCHESTRATORLead agent decomposes and delegates to specialistsHIERARCHICAL / HYBRID
REVIEWGenerate → critique → refine loopDEBATE / ENSEMBLE / EVALUATOR-OPT

Topology Selection

The deliberation engine (deliberation.ts) classifies every task and selects the optimal topology:

ComplexityDependenciesSelected Topology
TrivialNoneSINGLE
LowSequentialCHAIN
LowIndependentDISPATCH
MediumMixedORCHESTRATOR
HighMixedORCHESTRATOR
High-riskAnyREVIEW
CriticalAnyREVIEW
IterativeAnyREVIEW

Topology Details

SINGLE

One agent handles the entire task. Best for simple, well-scoped requests.

CHAIN

Agents execute in order, each building on the previous output via artifact references. Best for multi-step transformations.

DISPATCH

Independent subtasks run concurrently via sub-agents. Results are synthesized at the end. Best for parallelizable work.

ORCHESTRATOR

A lead agent decomposes the task and delegates subtasks to specialist agents, then synthesizes results. Adaptive rerouting allows mixed parallel/sequential execution.

REVIEW

Multiple agents independently produce solutions, then cross-validate and refine. Includes debate (cross-validation), ensemble (weighted voting), and evaluator-optimizer (generate-critique-refine) patterns.

Agent Scaling

The effortScaler.ts module scales the number of agents dynamically:

  • Simple tasks: 1 agent
  • Moderate tasks: 2–5 agents
  • Complex tasks: 5–10 agents
  • Research tasks: 10–20 agents

Agent Communication

Agents communicate through:

  • Message bus (messageBus.ts): Pub/sub for inter-agent and system events
  • Agent handoff (agentHandoff.ts): Direct agent-to-agent handoff with persistent inbox
  • Artifact system (artifactSystem.ts): Reference-based communication to prevent information loss
  • Three-layer memory: Shared working/episodic/long-term memory for context across agents

MIT Licensed — Built for multi-agent orchestration.