Your agent wanders when a goal has no plan

You gave your agent a multi-step goal and watched it try something plausible-looking on step one instead of figuring out the whole path first.

IntermediateRead 7 minDoing: about an hour, working from a tool loop you already have running7 steps

▶ Start

What you'll have at the end

A written plan of ordered subtasks, checked against the goal, that the loop executes one step at a time

Before you start

You need

A tool-calling loop that already sends the model one action at a time and stops once it says it's done, with nothing that writes out the whole task before the first tool call.

Not covered

This doesn't cover splitting the plan across two separate agents, or deciding whether a subtask needs a person's sign-off before it runs: both are their own problem.

Leans on

7 steps

Where to go next

Draft a handoff contract between two agents

for splitting the plan's subtasks across two separate agents instead of running every one in your own loop.

Should your agent ask before it deletes, sends, or pays?

for deciding whether a subtask needs a person's sign-off before the loop runs it.

Checked 15 Aug 2026

Part of the AI Agents cookbook