What is AI agent?
An AI agent is an LLM-driven system that can take multiple actions toward a goal — invoking tools, reading results, deciding next steps, and iterating — rather than producing a single chat response. The defining feature is autonomy: the agent decides what to do next based on what it has seen so far.
Also known as: LLM agent, autonomous agent, agentic AI
How agents differ from chat
A chat interaction is request → response. An agent interaction is goal → loop(decide → act → observe) → final result. The agent breaks a task into steps, picks the right tool for each step (search, code execution, file read, API call), executes it, reads the output, and adjusts the next step. The user gives one instruction and sees one final outcome, but many internal turns may have happened.
Common agent patterns
(1) ReAct (Reason → Act → Observe), the canonical loop. (2) Plan-and-execute: generate a multi-step plan first, then execute. (3) Tree-of-thoughts: explore multiple reasoning paths in parallel and pick the best. (4) Multi-agent orchestration: a planner agent dispatches subtasks to specialist agents (researcher, coder, reviewer). Most production agents use a flat ReAct loop with carefully designed tools rather than complex multi-agent topologies.
Tools an agent typically has
Web search, file reading/writing, shell command execution, HTTP requests, database queries, internal company APIs, and increasingly MCP servers (which expose standardized tools over a protocol). The agent's intelligence is bounded by its tools — give it good tools and clear descriptions, and it will outperform a chat-only model on complex tasks.
Where agents work and where they don't
Agents excel at well-defined multi-step tasks with verifiable outcomes: data extraction from many sources, code modifications across a repo, deep research with citations. They struggle with tasks that have ambiguous goals, with extremely long horizons (compounding errors), and with anything requiring real-world judgment. The frontier in 2026 is making agents reliable enough for end-user delegation rather than developer experimentation.