LLM-as-Agent
Strip an agent down and the surprise is how little of it is new: a language model, a loop, some tools, and somewhere to keep notes. LLM-as-agent is the name for that premise — the model supplies the judgment about what to do next, and ordinary software supplies the rest.
What the model is doing
In each turn the model receives the goal, the history of the run, and the schemas of the available tools — and emits one of two things: a tool call with arguments, or a final answer. The runtime executes the call, appends the result to the history, and asks again. The model is serving as the policy, the component that chooses the next action, and nothing in its training made it one. ReAct showed that prompting alone was enough to get the behaviour; native tool calling later made the output structured enough to rely on.
What the framing gets wrong
Reading it as "the model is the agent" leads straight into the characteristic failures. The model has no state between calls, so memory is whatever the scaffolding puts in the context window — and a goal that falls out of it is forgotten. It cannot verify a tool result, so a plausible wrong argument is executed as readily as a right one. It has no sense of having done enough, which is why step budgets and termination conditions live in the runtime, never in the prompt.
Why it matters
Agent quality is a property of the whole system, not the model inside it. The same model will solve a task with good tools, bounded context and a real stop condition, and spin for fifty steps without them — which is where the engineering is.
LLM-as-agent is the framing in which a language model serves as the decision-making core of an agent: given the goal, the run history and the available tool schemas, it chooses the next action, while a surrounding runtime executes tool calls, holds state and decides when to stop. It matters because agent quality follows from that scaffolding as much as from the model — the model selects actions but has no memory, verification or stopping condition of its own.
A clever person dropped into an unfamiliar office can still get work done — but only if someone gives them a desk, a phone, a filing cabinet and a note saying what the job is. LLM-as-agent is that idea applied to AI.
The AI model is the thinker. On its own it only answers questions, one at a time, and forgets everything in between. The surrounding software gives it the job description, the tools, a record of what has happened so far, and a rule about when to stop.
So when one of these assistants works well, it is rarely just a smarter model. It is a decent thinker with a well-organised desk.