"Agent" is the most overused word in AI right now. A lot of what gets called an agent is really a workflow — and that distinction matters, because the two have very different costs, failure modes, and debugging stories.
The difference in one line
- A workflow runs the LLM through steps you defined. The control flow is fixed.
- An agent lets the LLM decide the steps — which tools to call, in what order, and when it's done. The control flow is dynamic.
Both can use tools. Both can call the model multiple times. The difference is who's driving: your code, or the model.
When a workflow is the right call
Reach for a workflow when the task is predictable enough to map out in advance:
- Summarize a document, then extract fields from the summary.
- Classify a ticket, then route it based on the label.
- Generate a draft, then run it past a checker prompt.
Workflows are easier to test, cheaper to run, and far easier to debug — when something breaks, you know exactly which step failed.
When you actually need an agent
Agents earn their keep when you can't predict the steps ahead of time:
- The number of steps depends on what the model discovers along the way.
- The task needs open-ended tool use (search, then read, then search again).
- A human would also have to improvise rather than follow a checklist.
The tradeoff: agents are harder to evaluate, can loop or wander, and cost more because they make more model calls.
Rule of thumb: start with a workflow. Only reach for an agent when a fixed sequence genuinely can't express the task.
A quick decision table
| Question | Workflow | Agent |
|---|---|---|
| Are the steps known in advance? | Yes | No |
| Is cost/latency sensitive? | Yes | Less so |
| How easy is it to test? | Easy | Hard |
| Does it need to improvise? | No | Yes |
Most teams over-reach for agents and end up with something slow and unpredictable when a three-step workflow would have shipped. If you want to learn how to design both — and the evals that tell you which one is actually working — that's the core of our courses. New here? Start with what RAG is.