Give an Agent a Job It Can Finish

'Sort my entire life out' sounds like a tempting instruction for an AI agent. It's also an impossible brief. A better starting point might be: 'Find three suitable venues for Saturday and prepare a comparison.'

Agents work best when they know what a finished result should look like. A goal such as 'research venues' is open-ended; 'produce a shortlist with price, location and contact details' is a deliverable you can assess.

Suppose you're helping organise a community event. Set the date, budget, accessibility needs and how far people can travel. The agent can search, gather options and organise the findings without needing to invent your priorities.

You can also state a stopping point: don't contact venues, make payments or book anything. If information is missing, ask rather than assume. That makes the task clear without turning it into a hundred-line technical specification.

A small, well-bounded job is both easier to delegate and easier to learn from. Once you've seen what the agent can do reliably, you can decide whether expanding the task is worthwhile.

An instruction like 'Sort out my holiday' sounds convenient, but it isn't a workable job specification. Does sorting it out mean generating ideas, comparing prices, booking accommodation or contacting people? An agent needs a finish line and boundaries around what it is allowed to do.

Suppose your goal is to compare three possible weekends away. Give the system a budget range, travel distance, preferred accommodation and the date options. Then describe the output: a shortlist containing total estimated costs, travel information, the important drawbacks and links to the original options. That is something you can examine when the agent reports back.

Now add a stopping condition. Perhaps the system should return the best three suitable options or explain why it couldn't find three. Without this, an overly eager agent might continue searching, collecting ever more pages and consuming time or usage allowance without improving the decision. More activity is not automatically more value.

Consider how it should behave when constraints conflict. Imagine no accommodation matches your budget and accessibility requirements. Should the agent silently increase the budget, choose inaccessible options or ask which requirement can change? The third answer is usually the more responsible design. A hard requirement should not disappear because it makes the task awkward.

You can distinguish *requirements* from *preferences*. Step-free access might be essential, while a particular view or hotel style could be negotiable. Stating that difference helps the agent decide which alternatives to explore and which to reject. It also makes the eventual shortlist easier for you to assess.

A useful instruction describes not only the destination but the evidence needed along the way. Ask for the sources of prices and availability, the date those details were checked and any assumptions that affect the comparison. You can then decide whether to proceed. The system doesn't need to make the final purchase to provide a substantial benefit.

Think of this as delegating a bounded job to a junior assistant: you would explain what success looks like, which decisions they can make and when they should come back to you. The analogy isn't perfect, but it captures the operational discipline. Good delegation reduces ambiguity rather than simply demanding independence.

Let's examine what happens when an agent finishes early. Suppose it finds only two suitable weekends instead of three. It could fabricate a third option, quietly relax your requirements or report honestly that the search found two matches. Only the third respects the agreed outcome. A completed task can legitimately return 'no suitable result' when that is what the evidence supports.

You can also decide how much effort the task deserves. Perhaps fifteen minutes of searching is reasonable, but two hours isn't. A time or tool-use budget prevents the job from expanding indefinitely. That limit doesn't make the agent less ambitious; it tells the system how to balance its effort against the value of the result.

Before delegating, write one sentence explaining what you will do with the output. If you intend to make a booking decision, the shortlist must contain different information than if you're merely looking for inspiration. The clearer the next human decision, the easier it becomes to define what the agent should deliver.