Anthropic sells the tools for building AI agents. Their own engineering guidance tells developers to avoid building one whenever something simpler will do.

That reads like false modesty until you see the price of the autonomy.

A workflow is code where you fix the steps in advance: call the model, take the result, hand it to the next step you already wrote. An agent is code where the model itself decides what to do next, in a loop, picking its own tools and stopping when it decides it is done. One is an IKEA instruction sheet. The other is hiring a contractor and saying "renovate the kitchen, use your judgment."

The whole difference is who chooses the next step. And you pay for that.

Anthropic's team measured it on their own research system. A single agent uses roughly 4x the tokens (the chunks of text a model reads and writes, and what you get billed for) of a normal chat, and a setup with several agents working in parallel about 15x. Token spend alone explained 80% of the difference in how well it performed.

The bill is the visible cost. The quieter one: an agent takes a different route every run, so when it fails there is no fixed step to point at, and small errors compound down the loop.

Pay that premium when you genuinely cannot know the steps up front, like open-ended research that branches in ways nobody could list in advance. If you can already draw the flowchart, just write the flowchart.

Anyone can spin up an agent. Knowing when not to is the part that survives production.

Quick check before you scroll: Your team needs to pull three specific fields (name, date, amount) out of thousands of scanned receipts. Agent or workflow, and why?

Full breakdown + the answer: frankduah.me/learnings/2026-09-03-when-not-to-use-an-agent-simpler-is-often-better

New here? I post a bite-size AI / ML concept like this every day, follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH

#NOT #AI #LLM #AIAgents #MachineLearning