Beyond the Chatbox: What Is an AI Agent?

Transcript

SPEAKER_1: Alright, so I've been thinking about this a lot lately — the word 'agent' gets thrown around constantly now, but I'm not sure most people actually know what separates an AI agent from, say, just asking ChatGPT a question. SPEAKER_2: Right, and that distinction really matters. The key idea is that a standard chatbot is essentially reactive — you give it a prompt, it gives you a response. One turn, done. An agent is something fundamentally different. SPEAKER_1: So what makes it different? Is it just that it does more steps? SPEAKER_2: More steps, yes — but that's almost too simple. The real shift is in orientation. A chatbot is instruction-oriented. An agent is goal-oriented. You hand it a high-level objective, and it figures out the path to get there without you holding its hand at every turn. SPEAKER_1: Mm-hmm. So how does it actually figure out that path? SPEAKER_2: This is where the Perceive-Reason-Act loop comes in. Think of it as the engine. The agent first perceives information from its environment — that could be a webpage, a database, a file, whatever. Then it reasons about what to do next. Then it acts. And crucially, that action changes the environment, which feeds back into the next perception. SPEAKER_1: So it's a loop, not a line. It keeps cycling. SPEAKER_2: Exactly. And each cycle updates its understanding of where it is in the task. That's what allows it to handle multi-step processes — things like researching a topic, drafting a report, and then sending it — without a human prompting each individual step. SPEAKER_1: For example, suppose someone asks an agent to book the cheapest flight to Berlin next month. A chatbot tells you how to search. An agent actually searches, compares, and — depending on its permissions — books it. SPEAKER_2: That's a clean example. And notice what the agent needed to do that: it needed to use external tools. Search engines, APIs, maybe a calendar. The loop alone isn't enough — the agent has to be able to reach outside itself. SPEAKER_1: Wait — but that assumes the model already knows how to use those tools. How does that actually work? SPEAKER_2: Good catch. There's a mechanism called function calling — it gives the language model a structured way to request that an external program perform an operation. The model doesn't execute the code itself; it asks for the result, gets it back, and folds that into its next reasoning step. SPEAKER_1: So the LLM is really the reasoning core, not the whole system. SPEAKER_2: Precisely. The LLM handles reasoning. Around it, you have planning — breaking the goal into a sequence of actions. You have memory — both short-term, like what's happened in this task so far, and long-term, stored externally and retrieved when needed. And you have tool use. Strip any of those out and the agent gets significantly weaker. SPEAKER_1: [short pause] Now, there's a question that I think our listener — someone like Alina who's been thinking about alignment — would immediately raise here. Is the agent actually making choices? Or is it just a more elaborate deterministic script? SPEAKER_2: [inhale] Honestly, that's the sharpest question in this whole space. And the honest answer is: it's both, depending on how you define 'choice.' The agent is not conscious. It's not self-aware. A system can be fully agentic — planning, acting, recovering from errors — while being produced by a fixed model combined with prompts, tools, memory, and an execution controller. No magic required. SPEAKER_1: So the autonomy is real, but it's not the kind philosophers argue about. SPEAKER_2: Right. The autonomy is architectural. The agent can pursue a goal across many steps, adapt when something fails, and use tools to get information its training never contained. That's genuinely powerful — and genuinely risky, which is why safe deployment means controlling permissions, monitoring behavior, and always having a way to interrupt. SPEAKER_1: The takeaway for anyone trying to understand this space: the chatbot answers, the agent acts. And that loop — perceive, reason, act — is what makes the difference. SPEAKER_2: Well put. And in our next lecture, we move into the engine room — the specific components that make this autonomy work in practice. Memory architectures, planning strategies, how agents recover when things go sideways. The loop is the concept; next time we look at the machinery inside it.