Understand it in plain language

An AI agent can choose actions, call tools, inspect results, and continue toward a goal. A workflow follows a more predefined path; an agent gives the model more control over the next step. Names vary across products. Start with the simplest reliable design, and add autonomy only when it improves measured outcomes. Tool permissions, validation, stopping rules, and human review are part of the system.

How the AI approach works

An agent is a system that observes and acts toward an objective. In classical AI, agents include search-based planners and learned control policies. In current LLM applications, “agent” often means a model-driven loop that selects tools or next steps, observes results, and decides whether to continue. Tool use alone does not establish useful autonomy.

A workflow uses a largely predefined control path, such as retrieve → draft → validate. An agent delegates more control over the path to the model. Start with a workflow when the steps are known. An agent is worth considering when the useful next action depends on intermediate evidence and a fixed sequence becomes too restrictive.

A practical loop supplies the goal, allowed tools, relevant state, and stopping conditions; receives a proposed action; validates its arguments and authorization; executes an allowed call; returns the observation; and repeats within limits. Memory may mean short-lived task state or persisted records; neither should be confused with guaranteed correct long-term understanding.

Measure completed tasks, unwanted actions, recovery behavior, latency, and cost. Add budgets, timeouts, duplicate-action protection, and escalation. A model can propose a refund, but application code must decide whether the account, amount, and authorization permit it. Instructions inside retrieved material cannot grant new tool permissions.

Break the idea into steps

1. Give the agent only the tools and data needed for its task.
2. Validate tool arguments and results outside the model.
3. Limit steps and require review before important external changes.

Worked example: connect the method to the task

A course-support agent receives “I paid but cannot open Java.” It may inspect enrollment, check payment status, and read the access policy. If payment is unconfirmed, it should ask for the missing non-sensitive reference or escalate rather than invent success. If the verified condition permits access restoration, the application can expose a narrowly scoped restoration tool.

Test a timeout after an action, a repeated request, a misleading document, and an unavailable payment tool. A polished final message is insufficient if the underlying access change failed or happened twice.

Apply the idea: check your reasoning

Before you move on

• Explain the main idea without repeating the technical label.
• Identify the approach, its assumptions, and the evidence needed to trust the result.
• Complete the challenge, then compare your reasoning with the answer.

Sources & a deeper next step

The refresh uses primary papers and official documentation. These optional links are in English; you can understand this lesson without reading them first. Product versions and availability can change.

• Anthropic: building effective agents ↗
• NIST: Generative AI Profile ↗