A workflow follows a predetermined sequence of steps — the path is fixed by the developer, even if an LLM performs individual steps. An agent lets the LLM itself decide the sequence of steps dynamically, based on what happens along the way.
Side-by-Side
| Workflow | Agent | |
|---|---|---|
| Who decides the next step? | The developer, in code, ahead of time | The LLM, dynamically, at runtime |
| Predictability | High — the same input roughly follows the same path | Lower — the path can vary based on the model's decisions |
| Easier to test? | Yes — fixed steps are straightforward to unit test | Harder — requires evaluating decision quality across varied scenarios |
| Good for | Well-understood, repeatable processes | Tasks where the right steps can't be fully predetermined |
Example — Same Task, Both Approaches
# Workflow: fixed sequence, LLM used only for the "extract" and "summarize" steps
def process_document(doc):
text = extract_text(doc) # fixed step
entities = llm_extract_entities(text) # LLM used, but always runs, always this order
summary = llm_summarize(text) # fixed step
return {"entities": entities, "summary": summary}
# Agent: LLM decides what's needed based on the document
def process_document_agentically(doc, agent):
# Agent might decide: this doc needs OCR first, then extraction, then a
# follow-up entity lookup — or skip straight to summarization for a short doc.
# The exact steps are NOT fixed in code; the LLM chooses them at runtime.
return agent.run(goal="Process this document", input=doc)When to Prefer a Workflow
If you can write down the correct sequence of steps ahead of time and it rarely needs to change, a workflow is almost always the better engineering choice — more predictable, cheaper (fewer LLM calls making decisions), easier to debug, and easier to test exhaustively. Reach for an agent specifically when the right sequence of steps genuinely depends on information only available at runtime.
Practical Use Case
Generating a weekly sales report from a fixed data source is a workflow — the steps never change. Investigating "why did sales drop last week?" is a better fit for an agent — the right sub-questions to explore depend entirely on what the data shows at each step.
Common Mistakes
- Defaulting to an agent for tasks that are actually well-defined, fixed sequences — adds cost, latency, and unpredictability with no real benefit
- Building a rigid workflow for a task that genuinely needs dynamic branching, resulting in constant special-casing in code as new scenarios appear
Interview Relevance
"When would you choose a deterministic workflow over an agent?" is one of the most practically important agentic AI interview questions — a good answer focuses on predictability, testability, and whether the task's steps can genuinely be known ahead of time.
Practice Question
Decide workflow vs agent for: (1) converting an uploaded CSV to a fixed report format, (2) diagnosing why a specific customer's payment failed by checking multiple systems.
