Skip to the publication
AI Path Pro

AI PATH PRO / FIELDNOTES · EN

Goals, tools and decision boundaries

Write a job description for your agent.

Read the publication as a PDF or in the text edition below. You can also open and save the PDF separately.

2 pages · PDF
Download PDFOpen PDF file

Need help saving?

If the app does not offer saving, copy the file link and open it in Safari or Chrome.

Opening your publication…

AI PATH PRO / TEXT EDITION

Read it and put it to work.

AI Path Pro · 20 minutes

Define the goal your agent may pursue, the tools it may use and where it can choose the next step. State what counts as completion and when it must stop.

The method

  1. Give it a bounded goal

    Describe a result you can check. Preparing a sourced comparison of quotations is specific enough. Handling all purchasing does not yet tell you what to accept.

  2. Define the room to decide

    Which authorised source or tool may it choose next? Keep fixed business rules fixed. Instructions in a processed document cannot expand the agent's permissions.

  3. Design the stopping point

    What happens with missing critical data, conflicting sources, a tool failure or a reached time limit? Hand open questions and completed work back to the responsible person.

Preparing a supplier comparison

Illustrative example. The agent may read three supplied quotations and calculate from their line items. One quotation omits delivery costs. The agent cannot send external messages.

Goal: a comparison using equal quantities and a source for each item. Decision space: choose which quotation or calculation to inspect next. Gap handling: mark that quotation's total cost “cannot be determined”. Handoff: the comparison and a draft question requesting delivery costs. Supplier selection and external sending remain decisions for the designated person.

Your worksheet

Goal and completion

Which business task does it own? What checkable result allows it to finish?

Tools and decision space

What may it read, calculate or change? Which next steps may it choose independently?

Limits and handoff

What needs a human decision? What should it hand over, and to whom, after a gap, conflict or tool failure?

Trials and traceability

What should happen in normal and failing cases? Which sources and action statuses must remain traceable?

Review checks

  • Goal completion can be checked separately from confident wording.
  • Reading, modification and external sending permissions are defined separately.
  • The stopping condition has been tried with incomplete or invalid input.

Sketch the surrounding system

For agentic work, mark the trigger, the agent's task, validation and the human decision. Every handoff needs an input format, an output and an acceptance condition.

Take it into your own work.

Build my own starting plan