When Should You Use an Agent?

Before anyone builds an agent, walk through these questions. They separate genuine opportunities from expensive science projects.

• —Is the task repeated frequently? A one-off task rarely justifies an agent.
• —Does it require interpreting unstructured information — emails, documents, free-text requests?
• —Does it involve several steps that depend on each other?
• —Are the necessary tools and data actually available and permitted?
• —Can success be measured? If you cannot tell a good outcome from a bad one, you cannot evaluate the system.
• —What happens when the system makes a mistake? Who notices, and what does it cost?
• —Could ordinary automation solve it more simply?
• —What level of human approval is necessary, and who provides it?

Agents buy flexibility with complexity. Count the cost honestly:

• —Development cost: an agent needs instructions, tools, validation, and evaluation — more than a script.
• —Operating cost: every model call costs money or compute; multi-step tasks multiply it.
• —Latency: a loop of model calls is slower than a direct rule.
• —Reliability: model behavior varies; you must test failure cases, not just happy paths.
• —Maintenance: models, data, and tools change; behavior must be re-checked over time.
• —Privacy: whatever enters the context may leave your systems, depending on where the model runs.
• —Security: every tool is an attack surface; untrusted input can try to redirect the agent.
• —Human oversight: someone must review outcomes, especially early on. Budget for it.

For any scenario, your recommendation should be one of five: traditional automation, an AI model, an AI-assisted workflow (automation with a model step), an AI agent, or a human-led process. The right answer depends on the actual requirements, risks, and permissions — not on how impressive the technology sounds.

Automatically issuing refunds, for example, is usually a poor agent candidate: the action is sensitive, the rules can be made explicit, and mistakes cost money. Categorizing incoming emails, by contrast, is a classic fit: frequent, unstructured, multi-step, and easy to review.

Time to finish The First AI Assistant. In the exercise below you will produce a complete design specification: the problem, the user, the outcome, the context, the tools, the permission boundaries, the human approval points, the failure handling, and the success criteria.

The result is a learning exercise, not a validated production architecture — but every field you fill in is a question a real project must answer.

You can now explain what an agent is, tell it apart from models and automation, trace its execution cycle, configure its instructions, context, tools, and permissions, sketch its architecture, and evaluate whether a use case deserves one. That is the complete foundation — everything else in AI engineering builds on these six ideas.

When you are ready to go further, the guided project Build Your First AI Agent implements this flow in Python with a deterministic demo you can run without any paid account.