Lesson FP.1: What You're Producing

A complete design document for one agent, covering the decisions that are yours rather than engineering's. Whether or not you build it, the document is what a serious build review needs, and producing it will settle several arguments before they start.

Lesson FP.2: Part 1 - Scope, Autonomy and Tools

• One real task, with a step-by-step trace of how a person does it today including dead ends

• Your chosen autonomy rung, with a written justification for not choosing a higher one

• A tool inventory: name and four-sentence description for each, marked read or write

• Every write tool's worst wrong call, written in one line

• The confusable-tool check done, with descriptions rewritten where needed

• The minimum viable tool set identified, with what you cut and why

• The connections needed, and which already have a server versus which need building

Lesson FP.3: Part 2 - Permissions, Experience and Risk

• Every tool sorted into read, reversible write, or irreversible, with the outside-the-company rule applied

• Approval placement for each category, and the resulting approvals-per-person-per-day figure

• A designed approval screen for your riskiest action

• What flags a case as unusual and worth extra scrutiny

• The three display layers, with the actual running-commentary text written out

• The give-up message, written as the user will read it

• The four blast-radius questions answered, including how many times before anyone notices

• A one-page incident runbook, filed somewhere findable

Lesson FP.4: Part 3 - Evaluation, Cost and a Recommendation

• A 25-task test set from real examples, with good outcomes written by you

• At least two tasks where a damaging action would plausibly look helpful

• Ten runs scored by hand on success, path quality, constraint adherence and give-up quality

• Constraint breaches counted separately, any number above zero is a blocker

• A cost model with median, mean and the gap between them

• Per-task caps on steps, time and cost, with the expected hit rate

• The volume at which the economics stop working

• A one-page recommendation: ship, ship narrower, or don't, with the reasoning

Lesson FP.5: How to Know You're Done

Four honest questions. Could you have built this as a workflow instead, and can you say why you didn't? Have you written the worst wrong call for every write tool? Is your constraint breach count zero across your test runs? And would the person on call at 3am be able to stop this agent using only what you've written down? Four yeses and you have a design rather than an enthusiasm.