Skip to main content
designing your AI agent framework

What you will learn

Building an agent without understanding the work it needs to do is like hiring someone without writing a job description. You might get lucky. More likely, you will waste time fixing something that was never set up properly. This section gives you the method for designing agents that actually solve the right problem.

Map the work

Break down a current process into steps before deciding what to automate

What / When / How

Use a simple framework to define agent responsibilities clearly

Write a job description

Create a structured brief that translates directly into an agent build

Spot the blockers

Identify data access, integration, and permission issues before you start building

Start with the work, not the tool

The temptation is to jump straight into a platform and start building. That is the equivalent of hiring someone on day one without ever defining the role. The more valuable step is to slow down and look at the work itself. What are you doing today? Where does your time go? What is manual, repetitive, or frustrating?
This is not new thinking. It is the same process mapping you would do if you were restructuring a team or onboarding a new hire. The difference is that the new hire is an AI agent, and it needs even more clarity than a human, because it will do exactly what you tell it, no more and no less.

A real example: Pathfindr’s content pipeline

Internally at Pathfindr, the team faced this exact challenge. One team member was responsible for keeping the Academy Portal content up to date: updating blog posts when AI tools released new features, writing in the Pathfindr voice, structuring content into the right format, and publishing it. The manual process looked like this:
1

Check existing content

Browse what is already published to see what needs updating
2

Research updates

Search official documentation and release notes for the latest changes across ChatGPT, Claude, Copilot, and Gemini
3

Draft new content

Use AI to write a first version, applying the Pathfindr tone and structure
4

Review and refine

Add dynamic components, check formatting, and do a human review
5

Publish

Push the updated content live on the platform
Once mapped, the team could see which steps were candidates for automation. The research, drafting, and publishing steps became four sub-agents: a news researcher, a blog writer, an image manager, and a repository manager. The human review step stayed with the team member, because that is where their judgement adds the most value.
Map first. Automate second. The agents you build will only be as good as your understanding of the work they need to do.

The What / When / How framework

Once you have mapped the work, you need a way to translate each task into something an agent can execute. The framework is simple: What should the agent do? When should it trigger? How should it execute?

What

Define the task clearly. What is the agent responsible for delivering? What does the output look like?

When

Define the trigger. Is it time-based (every Monday at 9am)? Event-based (when a file is added to a folder)? On demand?

How

Define the execution. What tools does it need? What steps does it follow? Where does the output go?

Who reviews

Define the human checkpoint. Who reviews the output before it goes live? What level of oversight does this task need?

Example: Meeting preparation agent

Here is how the framework applies to a common leadership pain point: arriving at meetings unprepared.
Detail
WhatResearch attendees, pull related documents and recent emails, and send a briefing summary
When24 hours before each meeting in my calendar
HowCheck calendar for upcoming meetings. Look up each attendee on LinkedIn and in email history. Search Drive for related documents. Compile a briefing and send it to me via email.
Notice that each component is specific. Not “prepare me for meetings”, but exactly what to research, when to trigger, and how to deliver the output.

Writing the agent job description

When you put this together, you get something that looks like a job description for a new team member. This is deliberate. The clearer the brief, the better the agent performs. A strong agent job description includes:

Job title and purpose

One sentence on what this agent does and why it exists.

Responsibilities

The specific tasks it performs, each written in the what/when/how format. An agent might have one responsibility or five, just like a human role.

Data and access requirements

What tools, platforms, and data sources does the agent need to connect to? Gmail, Google Drive, LinkedIn, Slack, a CRM, a project management tool?

Potential blockers

What could prevent this from working? IT permissions, tool integrations that do not exist yet, data that lives in a system the agent cannot reach?
A single agent can have multiple skills or responsibilities. Your personal assistant agent might triage emails, send meeting briefings, and track project statuses: three separate responsibilities under one role. Each one follows the what/when/how framework independently.

The three-step evolution

The founder of Relay frames the agent-building journey in three stages. This is a useful way to think about where to start and where you are headed.
Start with tasks you are already doing that are boring, repetitive, or time-consuming. Filing documents into the right folder. Renaming files to match a convention. Sending RSVP reminders before meetings. These are low-risk, high-frequency tasks where agents deliver immediate value and you build trust in the system.
Move to tasks you are capable of doing but never have enough time for. Reading every competitor’s LinkedIn posts. Watching every relevant YouTube video in your industry. Scanning every blog post in your niche. You have the skill to do this analysis, you just do not have the hours. Agents give you that capacity back.
Finally, use agents for work that requires knowledge you do not personally hold. Financial modelling if you are not an accountant. Legal document review if you are not a lawyer. Technical analysis if you are not an engineer. The agent brings the subject matter expertise. You bring the business judgement about what to do with the results.
Most people should start at Step 1. Get comfortable. Build trust. Then expand.

Quick checkpoint

You are done with this section when you can:

Map a process

Break down one of your current workflows into clear, sequential steps

Apply What / When / How

Write at least one agent responsibility using the framework

Write the brief

Produce a job description for an agent that includes purpose, responsibilities, access needs, and blockers

Start at Step 1

Identify a mundane, repeatable task in your work that is a strong candidate for your first agent

Next: Building and Testing in Relay

Learn how to turn your agent design into a working workflow in Relay