Agent orchestration
Architectureagentic orchestration · orchestration layer
The layer that decides which agent acts, in what order, with which tools and which data — and that owns retries, escalation and handoff to a human. It is not the model. It is the system that coordinates several agents and tools to finish a job end to end.
Orchestration solves the problem that shows up the moment you have more than one agent: somebody has to divide the work. It covers task routing, state and memory across steps, tool invocation, error handling and retries, spend and time budgets, and the rules for escalating to a person. In Implementa's taxonomy, orchestration is the execution layer: governance says what is allowed, observability says what happened, orchestration makes it happen. The term has established public usage in agentic security and architecture writing — OWASP treats it under Planning & Autonomy and Tool Use — but framing it as a layer distinct from governance and observability is our own editorial formulation.
How it differs from
- Workflow automation (RPA)
- RPA runs fixed steps; agent orchestration picks the path at runtime.
- Agent framework
- The framework is the library you build with; orchestration is the operating policy you run on top of it.
Examples
- A support agent that triages, queries the CRM, drafts a reply and hands off to a human when confidence drops below threshold.
- A back-office pipeline where one agent extracts invoice data and another reconciles it against the ERP.
FAQ
- Do I need orchestration with a single agent?
- Yes, as soon as that agent calls external tools: someone has to set retry, timeout and escalation limits.
- Is it the same as multi-agent?
- No. Multi-agent is how many agents you have; orchestration is how they are coordinated.
Related terms
Sources & further reading
- OWASP — Agentic AI: Threats and Mitigations · 2025-02-17
- OWASP — Agent Control Standard (ACS) · 2026-09-01
OWASP GenAI Security Project — Agentic AI: Threats and Mitigations · 2025-02-17 · E3