A language model answers a prompt once. An agent uses a model in a loop: it observes, decides on an action, executes, and feeds the result back to decide again. That loop — not any framework — is what agent means, and it is the mental model everything else builds on.
The Core Loop
Every agent system reduces to the same cycle: the model receives a goal plus current context, chooses either a tool call or a final answer, the runtime executes any tool and appends the result, and the cycle repeats until the goal is met or a limit is hit. Frameworks change ergonomics, not this shape.
Autonomy Is a Dial Not a Switch
Between a chatbot and a fully autonomous system lies a spectrum: workflows with fixed steps and model-filled slots, routers that pick one branch, agents that choose their own action sequence, and multi-agent systems that delegate. More autonomy means more flexibility and strictly less predictability. Good engineering means choosing the lowest autonomy level that solves the problem.
Honest Expectations from Day One
Agents fail in ways ordinary software does not: they misread instructions, call the wrong tool with confident-looking arguments, loop without progress, or declare success falsely. Small per-step error rates compound, so a twenty-step task fails often even when each step is usually right. Design assuming failure, and give every agent step limits, timeouts, and a defined way to stop and ask.
Action Step
Pick a workflow you know well and sketch it at each autonomy level: fixed pipeline, router, and free agent. Write one sentence per level on what breaks when the model chooses wrongly, and note which level you would actually ship first.
Educational content only. The AI tooling landscape changes quickly — verify current library versions, APIs, and best practices in official documentation. This course does not certify anyone.