Agents are the exciting frontier of AI right now, so teams reach for them reflexively — wrapping a multi-step loop around problems a single prompt would solve. I think that's a mistake. An agent is the most powerful and the most unpredictable tool in the box, and most tasks don't need it.
Agents Trade Reliability for Autonomy
The whole point of an agent is that it decides its own steps. That autonomy is exactly what makes it unpredictable: more tool calls, more places to go wrong, higher latency, and higher cost. You take on all of that to buy flexibility. If the task doesn't need flexibility, you've bought unpredictability for nothing.
Most Tasks Are Actually Workflows
A huge share of "agent" use cases have a known sequence of steps. If you can draw the flowchart in advance, you don't need an agent deciding the steps at runtime — you need a fixed workflow calling the LLM at specific points:
# A workflow: predictable, testable, cheap summary = llm(f"Summarize: {doc}") category = llm(f"Categorize this summary: {summary}") save(category, summary)
That is faster, cheaper, and debuggable in a way an agent loop never is.
Use an Agent Only When the Path Is Unknown
The genuine case for an agent is when you can't know the steps ahead of time — the task branches based on what it discovers, and the number of steps varies by input. Open-ended research, interactive debugging, tasks that adapt to intermediate results. There, the loop earns its unpredictability:
Task: "Find out why this customer was double-charged."
(steps depend entirely on what each lookup reveals)
Where I Land
Default to the simplest thing that works: a single prompt, then a fixed workflow, and only then an agent. Reach for an agent when the path through the problem genuinely can't be planned in advance — and when it can, resist the temptation. A predictable workflow you can test beats an autonomous agent you have to babysit. Use the power where it pays for itself, not everywhere it fits.