What is an LLM agent?

The phrase is useful because it names the part that does the deciding. Older agents used rules, search algorithms or reinforcement learning policies. An LLM agent hands that job to a language model, which can read untidy inputs and write its own plans, but which can also be confidently wrong. Knowing that the decision-maker is an LLM tells you which failures to plan for: invented tool arguments, drift from the goal and sensitivity to wording.

Definitions vary, so check the source. The survey by Xi and colleagues (2023) describes LLM-based agents as artificial entities that sense their environment, make decisions and take actions, built around three parts it calls brain, perception and action. Anthropic does not treat “LLM agent” as a separate category; it describes agents as systems where an LLM directs its own process and tool use. On this site, LLM agent and AI agent mean the same thing unless a page says otherwise.

The best known pattern is ReAct, from a 2022 paper by Yao and colleagues. The model generates reasoning traces and task-specific actions in an interleaved manner: it writes a short thought, takes an action, reads what the action returned, and repeats. That lets the model adjust its plan when a tool returns something unexpected, rather than committing to a plan it wrote before seeing any data.

Modern model APIs build this loop in. The developer describes tools with a name, a description and a JSON schema; the model returns a structured tool call; the application runs it and sends back the result. The agent framework around the model handles memory, retries, step limits and logging.

Because the LLM is the decision-maker, the prompt and the tool descriptions are part of the program. Change a tool description and the agent may call it more or less often. Teams that run LLM agents in production treat prompts as versioned code and test them against a fixed set of tasks before each change.

A legal operations user asks: “Does our agreement with this vendor allow termination for convenience?” The LLM agent reasons that it needs the contract first and calls a document search tool with the vendor name. The search returns the master agreement and two amendments.

The agent reads the master agreement and finds a termination clause, then notices that the second amendment changes the notice period. It calls the search tool again for that amendment, reads the relevant section, and answers with both clauses quoted and cited. When it cannot find a clause, it says so instead of inferring one from typical contracts.

The behaviour that matters is the second search. A single retrieval would have missed the amendment. The model decided to look again because of what it read.

In Swfte Studio every agent is an LLM agent in this sense: its decisions come from a language model you choose. Because Studio reaches models through Swfte Connect, you can run different agents on different providers and change an agent’s model without rewriting the agent. Connect adds routing and fallback, budgets and an audit event stream for those model calls.

Studio records traces and token cost per step, which is how you see why an agent chose a tool. Golden sets and regression runs for agents are in development, not shipped, so for now keep your own fixed task set and run it before you change a prompt or a model.

• Swfte Studio
• Swfte Connect: One API in front of many model providers
• Models on the Swfte platform

Definitions on this page were read on the sources below on 7 October 2026. Where sources define the term differently, the page says so. The full glossary lists more terms.