
AI agent: a model that acts, not just answers
In short: An AI agent is an LLM running in a loop with tools and a goal: what it is, where you meet it and what it changes for engineering teams.
An AI agent is a language model that runs in a loop: it reads a goal, decides on an action, calls a tool such as search, code execution or an API, reads the result and decides again until the goal is met or it stops. Unlike a chatbot, the model chooses the next step. Anthropic's December 2024 guide describes agents as LLMs using tools in a loop.
AI agent in plain words
A chatbot is a well-read friend you ask questions. An agent is that friend with your laptop, your calendar and permission to click things. Useful, and also the reason you start caring about what it is allowed to touch.
How it works
Every agent has the same four parts: a model that reasons, a set of tools it can call, a loop that feeds tool results back in, and a stop condition such as a step limit or a completed goal. The surrounding code that runs this loop is usually called the harness.
Anthropic's guidance draws a useful line between workflows, where code decides the path, and agents, where the model decides it. The loop itself goes back to the ReAct paper from October 2022, which interleaved reasoning steps with actions. Most production systems are mostly workflow with a small agentic core.
Example
As an illustration, a support agent gets "refund order 8812". It calls get_order, sees the item was returned, calls issue_refund, then writes a confirmation. If the order is not found, it asks the customer instead of guessing. The whole task takes three tool calls and one decision, with a person as the fallback.
Where you'll meet it
- Coding tools such as Claude Code, OpenCode and Google Antigravity
- Customer support, sales development and IT helpdesk automation
- Research assistants that search, read and summarise
- Browser and computer-use products that operate software on screen
What it changes for you
For developers, the work shifts from writing every step to designing tools, permissions and checks. For tech leads, the question stops being "can the model do it" and becomes "how do we know when it did it wrong". Roles move towards evaluation, tool design and operations. The teams that do well treat an agent like a new hire with production access: scoped, logged and reviewed.
AI agent vs agentic workflow
An agentic workflow is a fixed sequence of model calls where your code picks the order. An AI agent picks its own next step. Start with a workflow; promote to an agent only where the path genuinely varies.
Common mistake
My view: most "agents" that fail in production failed at the tool layer, not the model. Vague tool names, no error messages and no stop condition make a strong model look weak. Fix the tools before you swap the model.
Related terms
Agent harness · Agent loop · Tool calling · Agentic workflow · Human-in-the-loop · Evals
From the archive: When you actually need an AI agent instead of a workflow · How to keep multi-step AI agents under control
An agent is only as good as the tools, permissions and stop condition you give it.


