The fastest way to build a disappointing AI agent is to begin with a vague, grand ambition.

  • “Handle my marketing for me.”
  • “Run all our customer service.”
  • “Become my chief of staff.”

Those are only aspirations, not workflows. An agent cannot reliably operate within a fog bank. A useful agent starts with clear goals:

  • Turn this meeting transcript into decisions, owners, deadlines, and unresolved questions
  • Review incoming enquiries, classify them, draft a response, and escalate the risky ones to my personal email
  • Gather material from these approved sources, create a first draft of a presentation, and flag any unsupported claims
  • Compare a published asset with an evaluation checklist, identify gaps, and ask for missing information

When I develop AI agents in New Zealand for clients, I use the same 4-step process I teach in the AI Agent Accelerator. It’s called the SAGE Framework, and it stands for: Scope, Automate, Generate, Evaluate.

1. Scope Agentic Tasks

Choose one workflow worth improving.

Projects fail when they aim too wide. By defining a narrow scope, you’ll learn how to identify agent-ready workflows, break them into steps, define success criteria, and decide where human review belongs.

A strong scope is narrow enough to test and valuable enough to matter.

2. Automate Your Workflows

Break the workflow into handovers.

Each tool connected to your systems creates another point of failure. Mapping each input, action, tool, handoff, and deliverable will create an automation map that allows you to debug individual steps.

Automation is a series of design decisions.

3. Generate System Prompts

Engineer agent instructions.

You want to give the agent enough context to produce good results without your active involvement. That means your instructions, skills, examples, and SOPs have to be fit for purpose.

Durable system prompts define success.

4. Evaluate Agent Performance

Test the output before you trust the workflow.

An agent becomes useful when you can tell the difference between a good result and a plausible-looking mistake. Building a simple evaluation can process to test outputs, spot failure points, reduce risk, and improve agent reliability over time.

Measure what matters.

This is the process I am teaching every Thursday, 10-11:30am, at EPIC Innovation for the next 8 weeks.

🏎️ “AI agents don’t have a speed limit, but the person managing them still does.” – Kieran Klaasen

Last Call for the AI Agent Accelerator

If you’d like to join me in this live, in-person 8-workshops series in downtown Christchurch, click here. (Newsletter subscribers get 25% off.)

If you’d like to have a conversation about hosting these workshops in-house with your team, click here.

🎬 [REPLAY] Human-First AI: Find the Work that Matters Most

[58:53]

👓 What I'm Reading

Upcoming Events:

Free webinar for members of the AI Coaching Academy:

Onboarding AI Agents: How to Delegate Workflows Friday, October 16 at 1:00 PM GMT+13

Onboarding AI Agents: How to Delegate Workflows

If the Internet gave everyone access to infinite information, AI has given us access to infinite skills. Managing these skills effectively will help you move beyond using AI as a research assistant, and into training an AI agent like a new hire on their first day of work.

RSVP

Premium webinar for members of the AI Campus:

Ai207 - AI for Learning Friday, October 23 at 11:00 AM GMT+13

Ai207 – AI for Learning

Turn Curiosity to Growth Build smarter learning habits with AI support. Learn how to personalise learning paths, reflect more effectively, and continuously upskill in a fast-changing environment. Create personalised learning systems using AI. Turn information into insight through reflection and synthesis.

RSVP