The Whole Course on One Sheet

Nodding means hearing the label without a usable explanation. Naming means saying what the term means and spotting misuse. Running means handing an agent a real job and judging the result. Directing means choosing suitable work, supplying requirements, checking the result and deciding what should change.

An AI agent receives a goal, uses the information and tools you permit, chooses some steps while it works and produces a result you can inspect. The word describes behaviour, not a product name. A product can call itself an agent while leaving important decisions or actions with you.

Look, Think, Do, Check. Look gathers the relevant information. Think selects or evaluates an action. Do carries out the permitted action. Check tests the result against the requirement and evidence.

A model processes supported inputs and produces outputs, including text and, where supported, images. Tools let the system read or change permitted resources. An API is a published way for software to request an operation. The environment is the information and systems around the task. The context window is the capacity for material the model can use at once. Memory or saved instructions can supply information to later runs. State records where the work has reached. A system prompt supplies standing product directions. Personalisation and saved task prompts can supply recurring preferences. A trigger starts the run.

A basic chat response answers the request you submit now. A fixed workflow follows a sequence someone designed. An agentic workflow lets the model choose some next actions from what the previous action reveals. A person, a schedule or an event can start a fixed or agentic workflow; the trigger is separate from the way the steps are chosen.

State the purpose, the boundary, the criteria in priority order and the output you need. Add examples when the task needs them. Ask for a useful explanation or calculation, but inspect the result against the evidence rather than treating extra words as proof.

Check whether the result is attached to the right task, answers your actual request and gives you something usable. Open the deciding sources, recompute important numbers, check the requested action and ask which parts were uncertain. Match checking effort to the consequence of an error.

Sending, spending, publishing and deleting can create consequences outside a draft. Keep approval or another control before those actions until the workflow is tested and its permitted scope is clear.

Ask the seller: can I watch the product do this with my own material? What information can it read, and what can it change? What decides its next step? What remains for me to do? What happens when it cannot finish? Where does a correction go, and how do you check the next result? What does the plan allow, and what stops usage from exceeding the budget? Which model is underneath, and what changes if the provider replaces it?

Training data shapes learned behaviour. Teleoperation means a person operates a machine from another location. A multi-agent system divides work among agents, while orchestration assigns work and routes results. Learning can augment rules-based, model-based, goal-based or utility-based designs; those labels describe decision behaviour, not a ranking.

Separate the plan price, usage or connected-service charges, preparation time, correction time, checking time and maintenance. A coverage percentage is useful only when its denominator is clear. Keep the reusable prompt, the dated output, the sources and the checking notes.

Start with one useful component whose input and output are clear. Test it, then connect a handoff when the next task has a defined input and check. Add coordination only when it solves a real repeated problem. Choose the smallest job you can inspect and improve.