Varun Vemulapalli

Day 86 of 100 · Phase 11: Capstone: build your agent

Designing your agent: use case, users, scope, success criteria

2 min readBy Varun Vemulapalli

Most agent projects fail before any code is written.

Not because the AI is weak. Because nobody wrote down what "done" means.

Today the capstone begins, and we start with the design.

Video explainer · 1:50
Key takeaway

Design an agent like a product: name the job, the users, the boundaries and the test for "done" before you write code.

Day 86 visual: Designing your agent: use case, users, scope, success criteria
The visual from the original post.

Meet MeetingMate, our running example for Days 86–96. Each post adds one piece.

The job: paste in meeting notes or a transcript. MeetingMate lists the action items as JSON (task, owner, due date), checks them against open tasks so nothing is duplicated, and drafts the follow-up email. A human approves before any task is created or any email is sent.

The concept

Before you build, answer four design questions:

  • Use case: what job, triggered when?
  • Users: who uses it, who is affected by it?
  • Scope: what it does, and what it will NOT do.
  • Success criteria: how you will measure "it works".

Walkthrough: designing MeetingMate

Our sample meeting (used until Day 96): a 30-minute product sync for a bank's mobile-app team.

  • Ana, product owner
  • Raj, engineer
  • Lena, compliance

It produces 3 action items:
1. Raj: fix the Android login timeout, by Friday.
2. Lena: review the new push-notification wording, by next Wednesday.
3. Ana: share the updated release plan with stakeholders, by Thursday.

Now the design card:

  • Use case: right after a meeting, turn notes into tracked tasks and a recap email in a few minutes.
  • Users: the organiser (Ana) runs it; attendees get tasks and email.
  • In scope: extract action items, check duplicates, draft the email.
  • Out of scope: joining calls, recording audio, sending without approval, setting priorities.
  • Success criteria (our targets, not industry figures): on a test set of sample transcripts (Day 91), find every real action item, invent none, get the owner right, flag duplicates, and leave nothing created or sent without a human "yes".

Build path: Python + Claude API tool use (Days 67–68), with a note on mapping it to Copilot Studio. Codex, GitHub Copilot or Claude Code can help write the code.

Where you see it at work

  • Product owners: this card is a mini product brief.
  • Business analysts: success criteria become acceptance criteria.
  • Banking: "out of scope" is where risk teams look first.

Try it today (10–15 min)

1. Copy the 5 headings: use case, users, in scope, out of scope, success criteria.
2. Fill them in for MeetingMate, or for the agent idea you picked on Day 85.
3. Write 3 things it must never do.
4. Turn each success criterion into a yes/no check you could test.
Building in Claude? Start here: https://platform.claude.com/docs/en/get-started

In one sentence

Design an agent like a product: name the job, the users, the boundaries and the test for "done" before you write code.

What would you put in MeetingMate's "out of scope" list? 👇