Kelly Mears

Multi-Agent Orchestration

Coordinating several model instances on one body of work, each with its own context.

Agents & Language Models2 min read231 words10 out · 10 in
also calledAgent orchestrationAgent fleet

Multi-agent orchestration runs several model instances against one problem, usually because the work decomposes or because independent perspectives are worth more than one long chain of reasoning. Each participant gets its own Context Window, which is the main structural benefit: work that would not fit in one conversation fits across several.

Common shapes:

  • Fan-out — split independent work across workers and merge results; see Fan-Out and Fan-In.
  • Pipeline — each item flows through successive stages without waiting for its siblings.
  • Panel — several workers attempt the same task independently and a judge selects or synthesises; see LLM-as-Judge.
  • Adversarial — an implementer and a fresh critic alternate; see Adversarial Review.

The hard problems are not about the model. They are the ordinary problems of concurrency, arriving without the usual guardrails: workers sharing a working directory overwrite each other; a generic scratch filename collides with a sibling's in-flight file; a bulk staging command sweeps another worker's uncommitted work into your commit; two workers independently invent conflicting names for the same shared concept. Isolation — a separate working copy per worker, explicit paths, unique filenames — is the answer, and it is not free. See Git Worktree and Race Condition.

Coordination has its own failure mode: a worker that "reports" into its own output rather than to the coordinator is invisible, and reads as failure when it merely finished quietly.

See also4

Hand-picked in the note itself — the neighbours worth reading next.

Linked from10

Notes elsewhere in the wiki that reach for this one.