The parts of an agent

Every AI agent, no matter how it's built, is made of the same handful of parts. Understanding these six pieces gives you the vocabulary and mental model for everything that follows — and they're all understandable without any technical background.

1. Instructions (the system prompt). These are the standing directions that define who the agent is and how it behaves — its role, its rules, its tone, and its limits. "You are a helpful research assistant. You always cite your sources. You never make up facts; if you're unsure, you say so." Instructions are the single highest-leverage thing you control — a well-instructed agent is dramatically better than a vague one. This is where most of your effort goes.

2. Tools (also called actions). Tools are what let an agent do things beyond talking: search the web, send an email, read a spreadsheet, look something up in a database, create a calendar event. Instructions make an agent smart; tools make it able to act. An agent with no tools can only produce words; an agent with tools can affect the real world — which is powerful and, as we'll see, risky. More tools mean more capability and more that can go wrong.

3. Knowledge (grounding in your data). This is how you give an agent your specific information — your documents, your FAQ, your policies — so it answers from real facts instead of guessing. The technical name is RAG (retrieval-augmented generation), but the plain idea is simple: the agent searches your documents, pulls out the relevant parts, and answers based on those rather than from vague memory. This is how uploaded files and knowledge bases work, and it's the main way to make an agent accurate about your world.

4. Triggers. A trigger is what starts the agent. It might be a person typing a request, a schedule ("every morning at 8"), or an event ("when a new email arrives," "when a form is submitted"). Simple chat-based agents are mostly started by a person; more advanced builders let agents run on schedules or in response to events.

5. Memory / context. This is what the agent remembers. Short-term memory is the current conversation (it can only hold so much at once — its "context"). Longer-term memory is notes or history it can save and look up later. Important honest note: memory is finite and imperfect — agents forget things and lose the thread on long tasks (Module 4).

6. Human-in-the-loop (approval). This is a required human check before the agent does anything consequential — sending, paying, deleting, publishing. It's the primary safety control for beginners, and you should treat it as non-negotiable for anything irreversible. A well-designed agent drafts and proposes; a human approves the important actions.

How they fit together. A useful agent = clear instructions + the right tools + relevant knowledge + a sensible trigger + enough memory + a human in the loop for the risky parts. Every agent you build in this course is just some combination of these six.

The mindset: every agent is built from six understandable parts. Instructions define who it is and how it behaves (your highest-leverage control); tools/actions let it do things beyond talking (making it able to act — powerful and risky); knowledge/RAG grounds it in your documents by searching and answering from them; triggers start it (a person, a schedule, or an event); memory is what it remembers (finite and imperfect, so it forgets on long tasks); and human-in-the-loop is the required approval before consequential actions — non-negotiable for anything irreversible. A useful agent is just a sensible combination of these six, and knowing them gives you the mental model for everything you'll build.