Agent Workflow — AI Glossary

Reference details

Related terms

• AI Agent
• Tool Use
• Prompt Chaining

How it works

1. A developer fixes the control flow in code — which steps exist, in what order, and where branches are allowed. The model does not decide the shape of the run.
2. Each step is a separate model call with its own prompt and its own narrow job, so a failure is attributable to one step rather than to the run as a whole.
3. State passes between steps explicitly, as arguments rather than as accumulated conversation, which is what makes a step replayable in isolation.
4. Where the pattern allows a choice — routing to one of several handlers, or looping until an evaluator passes — the model picks from a closed set the code defines, not from an open field of actions.
5. Failures are handled per step: retry this call, fall back to a cheaper model, or stop and escalate, decided by code rather than by the model's judgement.

Example: Why a fixed shape costs less to debug

A support workflow runs classify → retrieve → draft → check, four calls per ticket. A fully autonomous agent on the same task averages nine calls because it re-reads context and re-plans. At the same per-call cost the workflow is 4/9 of the spend, and when output is wrong there are four places to look instead of a nine-step trace that differs on every run. The trade is real in the other direction too: a ticket that needs a fifth step cannot get one, because no step in the graph produces it.

Commonly confused with

Often used interchangeably with 'agent', and the distinction is who owns control flow. In a workflow a developer fixes the steps and the model fills them in; in an agent the model decides what to do next and may take a path nobody enumerated. That makes workflows testable in the ordinary sense — the same input traverses the same nodes — and makes agents able to handle cases you did not foresee, at the cost of a trace that is different every time. Most production systems described as agents are workflows with one or two genuinely open decisions.

When to use it

Reach for it when:

• The task decomposes the same way every time, so the steps can be written down once.
• You need to attribute a bad output to a specific stage rather than to the system.
• Cost and latency have to be predictable, which requires a bounded number of model calls.

Reach for something else when:

• The right sequence depends on what earlier steps discover, so no fixed graph covers the input distribution.
• The action space is genuinely open-ended and enumerating branches would mean enumerating the problem.
• You are still learning the task shape — fixing a graph too early encodes a decomposition you will have to unpick.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Agent Workflow — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path