Designing agents — overview

What to decide before you build anything.

The questions that come before the code

Most agent projects that fail do not fail because the model was wrong. They fail because someone started wiring up tools before anyone answered the boring questions: what is this agent for, who owns it, and how will we know if it is working?

The Design pillar is those boring questions. It covers five decisions.

• Does this problem actually need an agent at all? A direct LLM call, a retrieval-augmented call, or a deterministic workflow may solve 60%+ of what you had in mind — with a fraction of the cost and risk.
• What kind of agent is the right fit? The axes are scope (narrow vs broad) and autonomy (assistive vs autonomous). Different quadrants carry very different governance burdens.
• What mental model do we use to reason about it? Managing a new employee — the four moves you already make as a manager map directly onto designing an agent.
• What is the agent's job, exactly? A Job-To-Be-Done statement, with an explicit NOT for list. Scope creep kills agent projects.
• How will we capture all of that in one page that the whole team can argue about? The Agent Canvas — eight questions, one sheet.

By the end of this module you will have drafted an Agent Canvas for a use case of your choice. You can use the interactive canvas as you go, or try a guided scenario.