Workflows vs. Agents: Choosing the Simplest Thing That Works
Most LLM features don't need an agent. A practical way to decide between a single call, a workflow, and a fully autonomous agent.
"Agent" has become the default word for anything that calls an LLM. That's a problem, because agents are the most expensive, least predictable way to use a model. Before reaching for one, it's worth walking up a ladder of complexity and stopping at the first rung that works.
Rung 1: A single call
Classification, extraction, summarization, and one-shot Q&A are a single request and a single response. You control the input, you validate the output, and you can test it like any other function.
Rung 2: A workflow
When a task has several steps but **you** know what they are, write them in code: retrieve, then draft, then check. The model fills in each step; your code owns the control flow. Workflows are easy to log, retry, and evaluate step by step.
Common shapes:
- **Prompt chaining** – the output of one call is the input to the next.
- **Routing** – a cheap classification call picks which specialized prompt handles the request.
- **Parallelization** – fan out independent sub-tasks, then merge.
- **Evaluator–optimizer** – one call drafts, another critiques, loop until it passes.
Rung 3: An agent
An agent is a loop where the **model** decides the next step: which tool to call, when to stop. Use one when the path genuinely can't be specified up front, such as debugging an unfamiliar codebase or researching an open question.
Four questions before you commit:
- **Complexity** – Is the task multi-step and hard to specify in advance?
- **Value** – Does the outcome justify higher cost and latency?
- **Viability** – Can the model actually do this kind of task well?
- **Cost of error** – Can mistakes be caught and rolled back (tests, review, sandboxing)?
If any answer is "no", step back down a rung.
The takeaway
Start with the simplest design, measure it, and add autonomy only where it measurably improves results. The best agent systems I've seen are mostly workflows with an agent loop in exactly the one place that needs it.
Newsletter
Get new posts by email
Practical notes on AI agents and system architecture. No spam, unsubscribe anytime.