AI AGENTS THAT WORKBuy the eBook

The free reader toolkit

Four questions to answer before you build an agent

Working tools built from the book's method. Take one real task from your organisation and walk it through in order, or jump to the question you are stuck on.

Runs entirely in your browser. Nothing you type is sent or stored.

Tool 1 of 4 · From Chapter 5 · The five-gate test2 min

Should this task be an agent?

Answer five questions about one specific task. The verdict tells you whether to build, decompose, fix the evidence first, or use plain automation instead.

Value: Does the task move a real business outcome?

A measurable outcome someone owns: cycle time, quality, risk or cost.

Boundary: Can trigger, owner and done be stated clearly?

You could write the first line of an agent contract today.

Judgement: Does reasoning or adaptation add value?

Inputs vary and rules alone fall short. If rules work, use rules.

Evidence: Can the decision and outcome be verified?

You have, or can build, examples with known right answers.

Control: Can consequences, tools and exceptions be bounded?

Permissions can be narrow, mistakes contained and exceptions routed.

Classify business-verification mismatches and recommend the next step

Promising

Promising. Tighten boundary and control first.

No gate fails outright, but the partial answers are where projects stall.

  • Boundary: write the trigger, owner and definition of done in one line each.
  • Control: narrow permissions and define where exceptions go.
Tool 2 of 4 · From Chapter 11 · Ownership, cost and value5 min

What will it really cost per good outcome?

The figures start from the book’s illustrative example. Replace them with your own. The model is rarely the biggest line.

p
%
min
%
£/hour
%
min
runs
£
min

Cost per good outcome

60.7p

  • Model and tools
  • Retries
  • Human review
  • Rework
  • Fixed costs
Cost per run
49.8p
Model and tools share of it
14%
Good outcomes a month
4,100
Agent cost a month
£2,488.75
Same outcomes done by hand
£17,220.00
Saving a month
£14,731.25

The biggest line is human review. Better evidence and fewer escalations will save more than a cheaper model. Move to sampled review only as the evals earn it. Fixed costs are set to zero: add platform, evals and run-team costs for the full picture.

Tool 3 of 4 · From Chapter 14 · Build your first agent3 min

Before you call it done

0 of 8 checked

Tool 4 of 4 · From Chapter 13 · The one-page Agent PRD1 hour

The Agent PRD template

The nine sections every agent needs agreed before anyone builds. Download it, fill it in with your team, and run the new-colleague test.

  1. 1. The decision this records
  2. 2. Outcome, cost and value
  3. 3. The prototype
  4. 4. Agent contract
  5. 5. Behaviour examples
  6. 6. Evals and graduation metric
  7. 7. Scope, non-goals and risks
  8. 8. Rollout, rollback and kill switch
  9. 9. Definition of done and change log

Get the book

The best agent is not the most autonomous one.

It is the one whose value, boundary, controls and evidence are clear.

Available now

Kindle eBook

Read tonight on any device. Searchable, with every canvas and checklist one tap away.

Buy on Amazon

Out 20 November 2026

Paperback

The desk copy for design reviews, with the workbook canvases ready to mark up.

Paperback out 20 November 2026

For organisations

Team copies and talks

Bulk copies for your team, or a talk or workshop on applying the method to your own agents.

Message Anuj on LinkedIn