All learning paths Learn See Try Make Write a task an agent can finish What this lesson is about Name the finished result, the limits, and the point at which it should come back and ask you.
A coaching centre asks you to set an agent on its admissions work this week. What do you write down first?
A Write down the topic: help with admissions this week, and let the agent work out the rest. B Write the steps you want it to take, and judge the run by whether it took them. C Name the finished result in a sentence a stranger could check.
Check my recall Why this matters Most first runs fail on the task, not on the agent. Anthropic's own engineering note says to include stopping conditions. One is a limit on how many times the loop may run. A task with no finished shape has no reason to end and no way to be checked.
Write an agent task with a checkable result, stated limits, and a rule for stopping.
Most first runs fail on the task, not on the agent. Anthropic's own engineering note says to include stopping conditions. One is a limit on how many times the loop may run. A task with no finished shape has no reason to end and no way to be checked.
Name the finished result in a sentence a stranger could check. State the limits first: what it may touch, what it may spend, when to ask. A task with no stopping rule will not stop, and that is the usual failure. Task Write an agent task for the admissions work at a small coaching centre.
Weak approach Writes: research the nearby colleges, find out what they charge, and help me with admissions this week. Nothing here says what a finished job looks like, or when to stop.
Stronger approach Writes: name the five nearest colleges under fifty kilometres, with the page that shows their charges. Send no mail. Stop at five, or at ten minutes.
Why the stronger approach works The weak task names no finished result and no stop. The strong one can be checked, and it can end.
Try a changed situation Your task already names its result. Now the agent will run on your shop's real order folder. What do you add before starting it?
Make something yourself Write one task for an agent with three parts: the finished result, the limits, and the stopping rule. Run it. Then write down what actually came back.
Lesson 2 of 9 on Running and checking AI agents next Watch one run from end to end Part 1 ends in a work sample you can send
Read the primary or official source: Anthropic · Building effective agents