How an AI Agent Works

Every tool-using agent runs the same basic cycle. Once you can picture it, most agent behavior — including most agent failures — starts to make sense.

1. 01Receive the request
2. 02Interpret the goal
3. 03Choose an action
4. 04Execute the tool
5. 05Observe the result
6. 06Continue or finish
7. 07Check the outcome

Walk through the cycle slowly. Notice how often the application — not the model — is the one acting.

• —Step 01 — Receive the request: the agent gets a goal from a user or an approved workflow. Example: “Find the total number of orders in the sample dataset.”
• —Step 02 — Interpret the goal: the model reads the request together with its instructions and the available context, and works out what is being asked.
• —Step 03 — Choose an action: if tools are available, the model may select one and propose the inputs — for example, a read-only counting tool over the orders table.
• —Step 04 — Execute the tool: the application validates the proposed action and runs it only if it is permitted. The model never executes anything itself.
• —Step 05 — Observe the result: the tool's output — say, 1,284 orders — is returned to the model as new context.
• —Step 06 — Continue or finish: the model decides whether another action is needed or it can now answer. The cycle may repeat, within a fixed turn limit.
• —Step 07 — Check the outcome: the application or a human verifies the result actually satisfies the original request.

A model has no operating-system access, no internet access, no database access, and no permission to perform actions. None of these come built in. Every capability an agent appears to have was explicitly provided by the surrounding application as a tool, and every permission was granted by the people who configured it.

This is the most common misunderstanding about agents. When an agent 'looks something up', the application looked it up. The model's job was deciding what to look up and explaining what came back.

Each step has a characteristic failure. Good agent design plans for all of them.

• —Interpretation: a vague request gets misread. Mitigation: clear instructions and clarifying questions.
• —Action choice: the model picks the wrong tool or wrong inputs. Mitigation: validation before execution.
• —Execution: the tool fails or times out. Mitigation: controlled errors returned to the model, not crashes.
• —Observation: the result is incomplete or misleading. Mitigation: the model is told what the tool could not do.
• —Continuation: the agent loops without converging. Mitigation: a hard limit on turns.
• —Outcome: the answer looks right but is not. Mitigation: evaluation against the original request, and human review for important tasks.

Take your First AI Assistant. A user asks: “Which of this month's documents are missing a signature?” Trace all seven steps for that request: what the assistant receives, how it interprets the goal, which tool it would need (a read-only document scan), what the result looks like, and who checks the final answer. You will configure exactly these pieces in the next lesson.

An agent works in a cycle: receive, interpret, choose, execute, observe, continue or finish, and check. The model interprets and proposes; the application validates, executes, and limits. Every step has a known failure mode, and every failure mode has a known mitigation.