Designing your first AI agent architecture

A practical walk through the decisions that matter before you write a line of agent code.

Before you build an AI agent, answer these four questions.

1. What is the unit of work?

An agent without a crisply defined task loops forever. Define the smallest complete outcome ('reply to this email', 'classify this ticket') and build for that.

2. Where does state live?

Agents fail when memory is an afterthought. Decide early: what does the agent remember between runs, and where?

3. What can it touch?

List every tool the agent can call and what the worst case of each call is. Your permission model is your safety model.

4. How do you observe it?

Log every decision with its context. When (not if) the agent does something strange, you need the trace.

Start with these, and the code becomes the easy part.