← AI Terminology

Agent

An agent is an AI system that perceives its environment, makes decisions, and takes actions — often in a loop — to accomplish a goal.

In the LLM context, an agent is a model given tools (web search, code execution, APIs) it can call autonomously across multiple steps.

Why It Matters in AI

Agents extend LLMs from single-turn answering to multi-step task completion — browsing the web, writing and running code, calling APIs, and revising their work based on results. They are the core architecture behind products like Devin (coding), Operator (web tasks), and Claude's computer use. Agents represent the shift from AI as a tool you query to AI as a collaborator that acts.

Key Points

Aspect	Description
Risk	Agents can take irreversible real-world actions — human-in-the-loop and sandboxing are critical
Action	Calls tools, writes code, sends messages, browses web, controls UI
Memory	Short-term (context window), long-term (vector DB or files), episodic (past action logs)
Planning	Chain-of-Thought, ReAct, or Tree-of-Thoughts to decide next steps
Frameworks	LangChain, LlamaIndex, AutoGen, CrewAI, Anthropic Agent SDK
Perception	Reads environment state — text, images, tool outputs, memory stores

Simple Analogy

A traditional chatbot is like a reference librarian — you ask, they answer. An agent is like a research assistant: given a goal, they go off, search databases, make calls, draft documents, iterate on feedback, and deliver a finished result.

Common Usage Examples

Claude using computer use to fill out a web form autonomously
OpenAI's Operator browsing the web and completing purchases on a user's behalf
AutoGen multi-agent conversation: planner agent + coder agent + critic agent collaborating
langchain.agents.AgentExecutor orchestrating tool calls in a loop
Devin (Cognition) — coding agent that writes, tests, and debugs entire software features

Summary

In short: An agent is an LLM that doesn't just answer questions — it takes actions, uses tools, and loops until the job is done.