What is agent orchestration?
Agent orchestration is the control logic that decides which agent, model call or tool runs next in an AI system, what context each receives and how outputs are combined, whether that logic lives in code or a model.
Every agent system has orchestration, even when nobody calls it that. In a single agent loop the model orchestrates its own tool calls. In a workflow, code fixes the order of steps. In a multi-agent system, some mix of the two decides which agent works on what.
The OpenAI Agents SDK documentation states the choice plainly: orchestration can be left to the model, which plans and picks its own steps, or written in code, which makes speed, cost and behaviour more predictable. The two can be mixed in one system.
Orchestration in code
Chaining
One agent's output becomes the next agent's input, such as research, outline, draft, critique and revise.
Classify then dispatch
An agent returns a structured category, and code picks the next agent from it.
Run in parallel
Independent agents run at once and code gathers their results.
Loop until it passes
A task agent and an evaluator agent run in a loop until the output meets the criteria or a cap is hit.
Where control should live
The labs differ here. Anthropic's guidance favours a simple loop in which the model drives, with scaffolding removed as models improve. Google's post on ADK 2.0 (July 2026) argues the other way for predictable processes: LLM loops used for routing are slow, costly and variable and tend to derail as context grows, so ADK 2.0 adds graph workflows that route in code and call a model only at the nodes that need judgement.
Google's illustrative refund example used about 50% fewer tokens per run (2,265 against 5,152) and about 20% less latency (5.7 against 7.2 seconds) as a workflow. The same post makes a security point: a manipulated model step cannot trigger an action that has no node or edge in the graph.
What the orchestration layer owns
State
What each agent has done, kept outside any one context window so a run can resume.
Context per step
Which part of the history, data and instructions each agent or call receives.
Limits
Iteration caps, token budgets, timeouts and retries for each step and for the whole run.
Approvals
Where a human must sign off before a risky action, and how the run pauses meanwhile.
Traces
A record of every step, so a failure can be found and replayed.
How Motiontic's engineers apply it
Motiontic's engineers put the predictable parts of a process in code and leave the model in charge only where the next step depends on judgement. Google's Cloud Architecture Center catalogue of patterns, from sequential and parallel to coordinator, hierarchical and human-in-the-loop, is a useful checklist, but the choice for each step comes from eval results on the client's real cases.
Code-driven compared with model-driven orchestration
| Code-driven | Model-driven | |
|---|---|---|
| Who picks the next step | Developer-written logic | The model, from instructions and tool descriptions |
| Cost and latency | Lower and predictable | Higher and variable |
| Handles surprises | Only paths someone coded | Can adapt to unplanned inputs |
| Testing | Unit tests on each branch | Evals over many runs |
| Typical failure | Rigid when the input does not fit | Skipped steps or loops as context grows |
Common questions
What is an agent orchestration framework?
A library that provides the control layer for agents: running steps, passing state, retrying, pausing for approval and tracing. LangGraph, Microsoft Agent Framework and Google's ADK all include one.
Is agent orchestration the same as multi-agent?
No. Orchestration applies to any agent system, including a single agent whose steps are partly fixed in code. Multi-agent systems simply have more to orchestrate.
Should an LLM or code decide the next step?
Code, where the steps are known and must be repeatable. The model, where the next step depends on judgement over unstructured input. Most production systems use both.
Related
Workflow vs Agent
A path fixed in code, compared with a model that chooses its own steps.
Agentic workflow
Model calls on a path set in code: predictable, cheaper and easier to test than a free-running agent.
Agent graph
Nodes for model calls and code, edges for what runs next, and state that flows through.
Multi-agent system
Several agents splitting one job, with some way of coordinating who does what.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.