Do you actually need an agent?

"Agent" is the word attached to almost every AI project proposal, which makes the vendor's own position on it striking. Anthropic recommends "finding the simplest solution possible, and only increasing complexity when needed" — adding that this "might mean not building agentic systems at all."

That is the company selling the agent tooling, telling you to check whether you need it.

The baseline you should beat

Before anything else, their stated default: "optimizing single LLM calls with retrieval and in-context examples is usually enough."

One call. Good context. A few examples. Most tasks that get scoped as agents land here once someone tries the simple version, and the simple version is cheaper to build, faster to run and far easier to debug when it misbehaves.

Build that first. If it clears your bar, you are finished.

Workflow or agent

Where one call is genuinely insufficient, there is a middle option people skip straight past.

Workflows run through predefined code paths. You decide the steps; the model does the parts that need language. Anthropic's case for them is "predictability and consistency for well-defined tasks".

Agents let the model direct its own process, choosing steps and tools as it goes.

The deciding question is whether you know the steps in advance. If you can write them down, encoding them as a workflow gets you a system that behaves the same way twice — which matters more than it sounds, given that the model does not behave the same way twice on its own.

Reach for an agent when the path genuinely cannot be known ahead of time: open-ended investigation, tasks whose next step depends on what the last one found.

The questions that settle it

Can I write down the steps? If yes, that is a workflow.

Does the next step depend on what the last one found? Only a genuine yes argues for an agent.

Have I built the single-call version and found it wanting? If not, you are choosing complexity without evidence, and the comparison you skipped is the one that would have decided it.

Can I check the output? An agent takes more actions with less supervision, so an unverifiable agent output is worse than an unverifiable single answer, not better.

If the answers do argue for an agent, the next decision is how to build one — the SDK, Managed Agents, or the raw API.

Try this

Take the thing you were going to build as an agent. Write the simplest single prompt that attempts it, with the relevant context pasted in, and run it on five real inputs.

Count how many are acceptable. That number is your baseline, and any agent you build has to beat it by enough to justify what it costs to run and maintain. Often it already clears the bar.

What goes wrong

Starting at the most complex option. Agent-first design skips the two cheaper things that usually work, and the debugging cost arrives later.

Confusing "uses tools" with "needs an agent". A single call that fetches a document and answers is a tool-using call inside a fixed path. That is a workflow, and it is fine.

Not measuring against the simple version. Without a baseline there is no way to know whether the complexity earned anything, so it never gets removed.

Ignoring context cost. An agent accumulates tool results across many turns, which is the context problem at speed. Anthropic's strategies for it — compaction, note-taking, sub-agents — are all work you take on by choosing an agent.

How to check it worked

Compare against your five-input baseline on the same inputs. If the agent is better by a margin that justifies the latency, the cost and the maintenance, the complexity earned its place. If it matches the simple version, you have learned something valuable and cheap: ship the simple one.