ML//agent

An agent is an AI system that repeatedly chooses actions from its observations in pursuit of a task, and it is the form a language model takes when a task needs more than one answer: searching, calling tools, editing files. In an LLM agent, the model proposes the next response or tool call and an agent harness executes the surrounding loop.

An agent is an AI system that repeatedly chooses actions from its observations in pursuit of a task, and it is the form a language model takes when a task needs more than one answer: searching, calling tools, editing files. In an LLM agent, the model proposes the next response or tool call and an agent harness executes the surrounding loop.

Tools expand the available action and observation spaces beyond text generation, and some actions exist only to acquire the evidence needed to choose later ones (active perception).

Planning may be explicit, implicit in repeated inference, encoded in a workflow, or distributed between model and harness.

State does not live automatically inside the model. Conversation history, files, memory, and environment state persist only because software stores and presents them again; the three levels at which an agent can know something (weights, context, external stores) are set out in agent memory.

The useful shift is from the model as an oracle to the model as a policy component. Agency belongs to the assembled system, never to one completion: the agentic system supplies the memory, permissions and environment that make the agent's trajectory possible, and agent autonomy measures how far that trajectory can continue before a human has to authorize the next step.

On the web the agent becomes a web agent (search, fetch, browse) and in practice a hybrid agent that routes between an index, direct navigation, browser automation and APIs by cost. It does not want to traverse the Internet; it wants to obtain something.