Agent Orchestration

Agent Orchestration

Also called: 智能体编排 · Orchestration layer

The control layer that decides which agent runs when, how outputs flow between agents, and when the whole process terminates.

OrchestratorRouteSequenceMergeStop
Schematic (simplified)

Questions orchestration answers

  • Routing: which agent or tool should handle this step?
  • Sequencing / parallelism: are steps serially dependent or can they run at once?
  • Merging: how are results from multiple branches combined, and how are conflicts resolved?
  • Termination: what counts as done, and what counts as a failure that should abort?

Orchestration can be hard-coded (deterministic, testable) or decided at runtime by a "coordinator" model (more flexible, less predictable).

Difference from a single agent loop

A single agent's loop only chooses "which tool next." The orchestration layer schedules across multiple agents or stages. Many production systems combine the two: deterministic orchestration on the outside for control, agent autonomy within bounded steps on the inside.

Common misconceptions

  • Orchestration does not have to be model-driven — a fixed state machine or DAG is often more reliable.
  • Routing every decision through one "coordinator" model makes the system hard to debug and reproduce.

Related terms

Sources

  1. Anthropic — Building effective agents (2024)
  2. LangGraph — Multi-agent concepts

Compiled 2026-08-29 · This glossary is compiled from public papers, official specifications, and common industry definitions, and is updated as the field evolves. Corrections welcome.