Multi-Agent Orchestration
Coordinating several model instances on one body of work, each with its own context.
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.
Subagent
A model instance spawned by another to handle a scoped task with its own context.
Agents & Language Models11 connections
Agentic Loop
The cycle of model turn, tool call, observation, repeat — the basic mechanism of an AI agent.
Agents & Language Models12 connections
Token Budget
The finite allowance of model usage a task may consume, and the design decisions it forces.
Agents & Language Models11 connections
Human in the Loop
Requiring a person's judgement at chosen points in an otherwise automated process.
Agents & Language Models9 connections
Linked from10
Notes elsewhere in the wiki that reach for this one.
- Adversarial ReviewAgents & Language Models
Reviewing work with an explicit mandate to refute it, from a perspective that did not produce it.
- Agentic LoopAgents & Language Models
The cycle of model turn, tool call, observation, repeat — the basic mechanism of an AI agent.
- Context WindowAgents & Language Models
The bounded span of tokens a model can attend to in a single request.
- Fan-Out and Fan-InAgents & Language Models
Splitting work across parallel workers and recombining their results.
- Git WorktreeVersion Control & Delivery
An additional working directory attached to one repository, each on its own branch.
- Human in the LoopAgents & Language Models
Requiring a person's judgement at chosen points in an otherwise automated process.
- LLM-as-JudgeAgents & Language Models
Using a language model to score another model's output against a rubric.
- Race ConditionTesting & Verification
A defect whose occurrence depends on the relative timing of concurrent operations.
- SubagentAgents & Language Models
A model instance spawned by another to handle a scoped task with its own context.
- Token BudgetAgents & Language Models
The finite allowance of model usage a task may consume, and the design decisions it forces.