I think we can get better performance out of an AI agent by giving it a workflow a human could follow. Agents are optimized to do human tasks the way humans do them.

Take searching for data. A folder with subfolders and files gives the agent somewhere sensible to start. It can inspect the names, choose a relevant folder, look at what is inside, and read the files it needs. Each discovery helps decide what to check next.

That is the useful part of the human workflow: a flow. Searching through that structure will give better results than searching a flat database and repeatedly guessing what to search for. The agent has a route to explore, and each step gives it context for the next one.

The names and organization are doing some of the work. Anthropic's context engineering guidance describes how paths, names, and timestamps provide these clues.

When creating a workflow for an agent, start with how a person would approach the task. Give it a clear starting point and a way to move from one useful piece of information to the next.

If the workflow you are creating mimics what a human would do, you will get better performance out of the box.