hawkes.aiKey ideas
- The one idea — An agent is just a model in a loop: it thinks, calls a tool, observes the result, and repeats — choosing each next step from what it just observed, until the goal is met.
- How it is shown — A single tool call (ask→run→return) wrapped into a spinning cycle — think, act, observe, repeat — that keeps going until the model emits a final answer.
- The trap to avoid — Thinking an agent is a fixed script a developer wired up — it isn't; the model decides each step live, from observations.
- What it sets up — If a chatbot and an agent run the same model, what separates them?
There's no special agent AI. An agent is a regular model wrapped in one loop: think, act, observe, repeat. That loop is the entire secret.
The one idea
An agent is just a model in a loop: it thinks, calls a tool, observes the result, and repeats — choosing each next step from what it just observed, until the goal is met.
Last group, the model could call one tool: it asks, the harness runs it, the result comes back. Powerful, but a single step. Now the leap: wrap that one action in a loop, and a passive text predictor becomes something that pursues a goal on its own. That's an agent. Here's the loop. The model looks at the goal and everything so far, and thinks: what's the next action? It emits a tool call. The harness runs it and drops the result back into context. Then the model looks again, with that result in hand, and decides the next step. The crucial word is repeat.
How it works — the demo
A single tool call (ask→run→return) wrapped into a spinning cycle — think, act, observe, repeat — that keeps going until the model emits a final answer.
It's not one call, but a cycle: search, read the result, search again, calculate, write. Each pass, the model picks what to do next from what it just learned. The loop runs until the model decides the goal is met, then gives a final answer. Notice what this means: the model isn't following a fixed script. Nobody wrote step one, step two. At every turn it decides the next action from what it just observed. That's why agents feel adaptive, reacting to a failed search or an error, and why they're unpredictable. This think-act-observe cycle is the entire engine. Everything about agents, planning, self-correction, tools, getting stuck, is a variation on this one loop. A model, a set of tools, a goal, and a loop wrapping them.
The trap to avoid
Thinking an agent is a fixed script a developer wired up — it isn't; the model decides each step live, from observations.
Why it matters — and what’s next
If a chatbot and an agent run the same model, what separates them?
That's the whole machine. The trap: imagining an agent is a fixed workflow a developer wired up. It isn't. The model decides each step live, inside the loop, from what it observed. That's its power and its danger. You set the goal and the tools; the model chooses the path. So an agent is just a model in a loop: think, act, observe, repeat, until done. That idea is the foundation of everything ahead. But a chatbot and an agent can run the same model, so what separates them? Next: chatbot versus agent.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
Last group, the model could call one tool: it asks, the harness runs it, the result comes back. Powerful, but a single step. Now the leap: wrap that one action in a loop, and a passive text predictor becomes something that pursues a goal on its own. That's an agent.
Here's the loop. The model looks at the goal and everything so far, and thinks: what's the next action? It emits a tool call. The harness runs it and drops the result back into context. Then the model looks again, with that result in hand, and decides the next step.
The crucial word is repeat. It's not one call, but a cycle: search, read the result, search again, calculate, write. Each pass, the model picks what to do next from what it just learned. The loop runs until the model decides the goal is met, then gives a final answer.
Notice what this means: the model isn't following a fixed script. Nobody wrote step one, step two. At every turn it decides the next action from what it just observed. That's why agents feel adaptive, reacting to a failed search or an error, and why they're unpredictable.
This think-act-observe cycle is the entire engine. Everything about agents, planning, self-correction, tools, getting stuck, is a variation on this one loop. A model, a set of tools, a goal, and a loop wrapping them. That's the whole machine.
The trap: imagining an agent is a fixed workflow a developer wired up. It isn't. The model decides each step live, inside the loop, from what it observed. That's its power and its danger. You set the goal and the tools; the model chooses the path.
So an agent is just a model in a loop: think, act, observe, repeat, until done. That idea is the foundation of everything ahead. But a chatbot and an agent can run the same model, so what separates them? Next: chatbot versus agent.