Glossary
What is LLM agent?
Also known as: large language model agent, AI agent
An LLM agent is software that uses a large language model as its decision-making layer to perform multi-step tasks autonomously. Unlike a single LLM call (which produces one response), an agent runs in a loop: it observes the current state, chooses an action (often by calling external tools or APIs), observes the result, and repeats — until the task is complete. This pattern is the foundation of modern AI voice agents, autonomous coding assistants, and many "AI workflow" products.
The agent loop
A typical agent loop has four steps: observe (read the current state — the conversation so far, the user input, the result of the last action), think (the LLM reasons about what to do next, often producing a chain of thought), act (call a tool, send a response, update state), and repeat. The loop terminates when the LLM decides the task is done or when an external condition (max iterations, time budget) is hit.
In voice agents, the "act" step is usually one of: respond with speech, transfer the call, capture data, or call a tool like book_consultation(). The loop runs once per conversational turn.
Tools and function calling
The defining feature of an LLM agent (versus a plain LLM) is its ability to call external tools. The agent is given a list of available functions (e.g., create_clio_matter, send_text_message, check_calendar_availability) along with their input schemas. The LLM decides when to call which function based on the conversation. Modern LLMs from OpenAI, Anthropic, and Google all expose structured function-calling APIs that make this reliable.
Why agents matter for production work
Most real-world tasks aren’t one-shot. Booking a consultation requires checking availability, asking the caller, confirming the slot, sending a confirmation, and creating a calendar event. Qualifying a legal intake requires asking questions in the right order based on previous answers. Resolving a customer support issue requires looking up the customer record, checking the relevant policy, and sometimes escalating. All of these are agent loops, not single LLM calls.
Limits of LLM agents
LLM agents have meaningful failure modes. They can loop forever on ambiguous tasks. They can call tools incorrectly when prompted weakly. They sometimes hallucinate data when they should ask for it. Production agents include guardrails: max iteration counts, schema validation on tool calls, fallback to human escalation, and explicit termination conditions.
Voice agents in particular need turn-taking and interruption handling on top of the basic agent loop — the LLM has to know when to stop talking and listen.