LLM Agent

An LLM agent is a model that perceives an environment and acts on it through a tool inventory, with the model itself serving as the planner.

An agent is anything that can perceive its environment and act upon that environment. In foundation-model systems, the model is the "brain": it processes the task and environment feedback, plans a sequence of actions, and decides when the task is done.

Two design axes define an agent:

  1. Its environment — the use-case world it operates in (a codebase, the web, a database, a kitchen, a road system).
  2. Its set of actions, augmented by its tool inventory.

Environment and tools are interdependent: the environment determines which tools are possible, while the tool inventory restricts which environments the agent can effectively operate in. Familiar products already fit this frame — chat assistants with search and code execution, or RAG systems whose retrievers and SQL executors are tools.

Agent success depends on tool inventory quality and planner strength. Agents typically need stronger models than non-agentic flows because compound mistakes accumulate across steps and tool access raises the stakes of each wrong action. Teaching which, when, what, and how for tools can be prompted (function calling, ReAct) or learned via Toolformer-style self-supervision — either way, the harness must execute calls safely and feed results back as observations. That harness is a design surface in its own right: an agent's behaviour comes from the model plus its harness, not the model alone.

Sources

  • AI Engineering (Chip Huyen), ch. 6
  • Toolformer (Schick et al.), pp. 1–17