Multi-agent orchestration

Real workflows rarely fit in a single agent. ALDO AI ships six composition patterns that cover the bulk of multi-agent shapes you'll need; pick one (or layer them) on a single composite spec.

#Strategies

#Pipeline (sequential)

Each subagent receives the previous one's output. Use when you have a strict order: extract → summarise → format.

composite:
  strategy: sequential
  subagents:
    - extractor
    - summariser
    - formatter

#Parallel

Subagents run concurrently against the same input. The aggregator combines their outputs. Use when each subagent contributes a different facet (citations, sentiment, key entities).

composite:
  strategy: parallel
  subagents:
    - citations
    - sentiment
    - entities
  aggregator: combine-facets

#Supervisor

A supervisor agent decides which subagent runs next and reads results to decide if it's done. Use for open-ended planning.

#Router

A lightweight classifier picks one subagent to run. Use for triage (intent classification → specialised handler).

#Debate

Two or more subagents argue; an aggregator picks a winner or synthesises a consensus. Use for high-stakes generation (legal, medical).

composite:
  strategy: debate
  subagents:
    - position-a
    - position-b
  aggregator: judge

#Subscription

A subagent reacts to events from another agent's stream — useful for "watcher" agents that monitor a long-running run and intervene when a condition is met.

#Cross-field rules

The registry enforces:

• aggregator only applies to parallel and debate.
• iterative (a sub-mode of supervisor) requires exactly one subagent.
• Privacy tier cascades: a parent's tier is the floor for all children.

#Spawning rules

Spawning is narrow: only supervisors can spawn, and only the roles listed under spawn.allowed. Reviewers and auditors never spawn. This keeps the agent tree shape predictable and easy to reason about during replay.

#Related

• Replayability
• Building an agent