Orchestrator

The orchestrator is the part of an agent system that is not a model. It holds the workflow's state, calls agents and tools in order, passes results between them, counts steps and spend, applies timeouts and retries, and hands off to a person when a rule says so. It is where policy gates, step budgets and audit logging live, because those must not depend on a model remembering to apply them.

Orchestrators come in three flavours. Framework-based, using a library such as LangGraph, which gives you graph-structured control flow quickly. Custom, written as ordinary application code, which is often simpler and easier to test for a well-defined workflow. And workflow-engine based, using a durable execution system such as Temporal, which is the right choice when steps take hours, need retries across failures, or must survive a restart.

An orchestrator is not itself an agent, though a planner agent may advise it. The distinction matters for safety: the orchestrator enforces the boundaries; the models operate inside them. A design where a model can change its own step budget or skip a gate has confused the two.

• CTO / Head of Engineering: the orchestrator choice decides testability, durability and how hard it is to change the workflow later.
• CISO: it is the enforcement point for permissions, step limits and logging, which should be code, not prompt instructions.
• Operations head: retries, timeouts and escalation rules are what make an agent workflow behave predictably during outages and peak load.
• CFO: step and token budgets in the orchestrator are the practical cap on runaway inference costs.

Without an orchestrator, control lives in prompts, and prompts are not a reliable place for limits, retries or audit logs. The orchestrator exists to put deterministic control around non-deterministic models: to guarantee that a gate is checked, a step count is enforced and every action is recorded, regardless of what the model says. The trade-off is framework choice: heavy frameworks add abstraction that can be hard to debug, while custom code needs discipline to stay maintainable. The right choice depends on how long-running and how variable the workflow is, not on which library is popular.

• A support-agent system where the orchestrator routes tickets to specialist agents and escalates after two failed resolution attempts.
• A lending pipeline where the orchestrator holds a loan file's state across document collection, verification and underwriting over several days.
• A voice-agent stack where the orchestrator manages turn-taking, tool latency budgets and hand-off to a human line.
• A logistics exception workflow where the orchestrator retries customer contact across channels and updates the TMS on completion.
• A course-generation pipeline where the orchestrator sequences drafting, curriculum checks and editorial review with human sign-off.

Tool / technologyA software component, whether a framework, a workflow engine or custom code. Eazyware selects and builds the orchestration layer as part of multi-agent systems, choosing between LangGraph, durable workflow engines and plain code based on the workflow's shape.

Eazyware service that covers it: Multi-Agent Systems & Workflow Orchestration. Starting prices are on the pricing page.