An AI agent is an LLM running in a loop with tools and a goal. A chatbot answers one message at a time; an agent takes a task — “fix this bug”, “book me a flight”, “research this market” — then plans, acts, reads what happened and acts again, without a human approving every step. The model is the same technology; the loop and the tools are what make it an agent.
How the loop works
Each cycle has three beats. The model reasons about the current state (“tests fail with a type error in auth.ts”), calls a tool — running a shell command, editing a file, searching the web, clicking a browser — and then reads the result, which is appended to its context window as the input for the next cycle. Repeat until the goal is met, the model gives up, or the context runs out. That last clause matters: an agent’s working memory is its context window, and long tasks live or die on how well the harness compacts old history.
Why agents took over in 2025–26
Two things converged. Models got dramatically better at tool calling — emitting correctly-formatted actions instead of describing them — and labs started training directly on long-horizon agentic tasks rather than single-turn chat. The visible results: coding agents like Claude Code and Codex that run for hours on real repositories, computer-use agents that operate desktop apps from screenshots, and deep-research agents that browse for twenty minutes and return a cited report. Frontier models like Claude Opus 4.8 and GPT-5.6 now compete primarily on agent benchmarks, not chat quality — our best LLMs for agents ranking tracks that race.
Where agents still fail
Errors compound. A model that gets one step right 95% of the time finishes a 20-step task less than 40% of the time, so long tasks need checkpoints, tests or human review at the boundaries. Agents also inherit every LLM weakness — hallucination becomes acting on an invented fact, which is worse than stating it — and they add a new one: prompt injection, where text the agent reads (a web page, an email) contains instructions it mistakes for yours. The practical rule in 2026: give an agent real autonomy only inside a sandbox where failure is cheap, and keep a human on the irreversible steps — payments, deletes, sends.