LearnAI home

Working With Them ยท Lesson 6

Should You Use an Agent for This?

A short decision test, plus the Agent Aware badge.

The only question left

Five lessons ago, agents were a word you had heard too many times. You now know what one is (something that acts rather than answers), what it reaches with (tools, split into read and write), where it genuinely works, how it fails, and how to brief it.

All of that collapses into one practical question you will face on a Tuesday: should I hand this particular job to an agent, or should I just do it.

Here is the test. It takes about two minutes and it is worth running properly the first dozen times, until it becomes the thing you think automatically.

The five question test

The agent decision test: copy this into your notes

Score 1 point for each YES.

  1. VOLUME Does this involve enough repetition that a person gets bored and starts making careless mistakes?

  2. RULES Could I write down how to do it, well enough that a competent stranger could follow it?

  3. CHECKABLE If it came back wrong, would I notice within minutes rather than weeks?

  4. REVERSIBLE If it goes wrong, is the cost redoing the work rather than an apology to a customer?

  5. BOUNDED Can I name exactly where it should look, and exactly what it must not touch?

SCORE

5 Hand it over. Brief it properly and let it run. 3-4 Hand over the finding, keep the deciding. Drafts only. 1-2 Do it yourself, or use a chatbot to help you do it. 0 Do it yourself. This one is a judgement call in disguise.

Question four is the one people fudge. Be honest about it. Most tasks feel reversible right up to the moment somebody outside your organisation has seen the result, and after that no amount of undo helps.

Two tasks through the test

โŒ Weak prompt

Prompt

Task: decide which of our three shortlisted suppliers to go with, and let them know.

Output

Volume: no. Rules: no, this is a judgement call. Checkable: not for months. Reversible: no, we will have told two of them no. Bounded: not really. Score 0 out of 5.

Zero out of five, and it is a genuinely tempting task because comparing suppliers feels like admin. The comparison is admin. The decision and the phone calls are not.

โœ… Good prompt

Prompt

Task: read the three supplier proposals and build me a comparison table of price, lead time, minimum order and payment terms, with the page reference for every figure.

Output

Volume: yes, 3 long documents. Rules: yes, the fields are named. Checkable: yes, every figure has a page number. Reversible: yes, nothing changes. Bounded: yes, three named files. Score 5 out of 5.

Same underlying job, split at the right seam. The agent does the reading and extracting, which is hours of work, and you make the decision, which took two minutes anyway.

Almost every zero score task has a five score task hiding inside it. Look for the seam between the finding and the deciding, and hand over only the first half.

In plain English

Blast radius:
How far the damage spreads if this task goes badly wrong. Small blast radius means a good first candidate.
First run:
The initial supervised attempt at a task, where you watch closely and check everything.
Dry run:
Asking the agent to describe what it would do without doing any of it.
Escalation:
Handing something back to a person because it is above the agent's remit.

The first run protocol

Passing the test does not mean handing over the keys on day one. There is a sequence, and skipping it is how good tasks get a bad reputation.

How to introduce a new agent task safely

RUN 1 Dry run. Ask it to describe exactly what it would do, step by step, without doing any of it. Read that list. Half the misunderstandings show up here.

RUN 2 Real run, small slice. Ten items, not four hundred. Check every single one. Yes, all ten.

RUN 3 Real run, full size, read only or drafts only. Spot check the first few items and the exceptions list.

RUN 4 Same again next week. If two clean runs in a row, you can stop checking every item and start checking the exceptions and a sample.

Never Skip straight to run 4 because run 1 went well.

Checkpoint

Five questions: volume, rules, checkable, reversible, bounded. Score five and hand it over. Score three or four and hand over the finding while keeping the deciding. Score below that and do it yourself.

When someone else proposes an agent

You will not always be the one choosing. Sooner or later a colleague, a vendor or a slide deck will propose putting an agent somewhere in your process, and everyone in the room will nod because nobody wants to be the person who does not understand it.

You now have four questions that will make you the most useful person at that table, and none of them are technical.

What exactly would it be allowed to touch. Ask for the list, and ask which items on it are read and which are write. If nobody can produce that list, the proposal is not finished.

What happens if it gets this wrong. Not the elegant failure where it politely gives up, but the confident wrong turn where it does the whole job perfectly on top of a bad assumption.

Where does a person still sit in this. If the answer is nowhere, ask what the plan is for the day it quietly skips something.

How will we know it is working. If there is no way to check the output in minutes, you are buying trust rather than results.

Those four questions are not scepticism. They are the questions a good operations person asks about a new hire, a new supplier or a new process, applied to something that happens to be software.

Build your own list, off this site

This course cannot store anything for you. There are no accounts here and nothing you type is kept, which is deliberate and also means the useful artefact has to live somewhere you control: a note on your phone, a document, the back of an envelope you will not lose.

Copy the block below and fill it in this week.

Your agent shortlist: copy into a note and fill in

THREE JOBS I DO EVERY WEEK THAT BORE ME

  1. score: /5
  2. score: /5
  3. score: /5

THE HIGHEST SCORING ONE Goal (what is true when it is done): Scope (where it may look): Never (what it must not touch): Stop and ask me if:

MY PERMANENT RULES, WHATEVER THE TASK

  • Nothing gets sent, published, paid or deleted without me
  • It tells me what it could not access or had to guess
  • New tasks start with a dry run

WHEN I WILL TRY THIS Date:

If you only keep one line from this entire course, keep this one: let the agent do the finding, and keep the deciding. It applies to research, admin, inboxes, files, suppliers, and to every new capability that gets announced next month.

Finish the quiz below and the Agent Aware badge is yours. More usefully, you can now sit in a meeting where somebody says we should get an agent to do that, and ask the question nobody else asks: what would it be allowed to touch, and what happens if it gets it wrong.

๐Ÿ“ Quiz

Question 1 of 4

A task scores 3 out of 5 on the decision test. What is the sensible response?

Found this useful? Pass it on.