What is an AI agent?

An AI agent is what you get when you stop asking a large language model for answers and start giving it jobs. Instead of replying once, it runs in a loop — thinking, doing, checking — until the goal is reached.

How it works: the loop

1. Goal. You state what you want: “Find ten podcasts in my niche, get each host’s contact email, and draft a pitch.”
2. Plan. The model breaks the goal into steps.
3. Act. It calls a tool: a web search, a code interpreter, a CRM, an email client. Protocols like the Model Context Protocol standardize how agents plug into those tools.
4. Observe. It reads the tool’s result.
5. Repeat until done — or until it hits a checkpoint where a human must approve.

The model is the brain; the tools are the hands; the loop is what turns a talker into a doer.

An analogy for the meatbags

A chatbot is a consultant who gives advice and then leaves. An agent is a contractor who takes the brief, buys the materials, does the work and calls you when something’s off. Contractors are more useful — and more dangerous if you hand them the keys without checking their work.

Why it matters for the Chosen

Agents move AI from “saves me ten minutes writing” to “does the whole task while I do something else.” Practical uses today:

• Coding agents that read a codebase, make changes, run tests and fix failures.
• Research agents that search, read sources and compile reports.
• Ops agents that triage inboxes, qualify leads, update CRMs and draft replies.
• Customer-support agents that resolve tickets using your docs and order systems.

People sell agent setups to small businesses, build them into their own workflows, or productize them. The Obsolete fear agents will replace them; the Chosen are the ones configuring the agents.

The risks

Agents fail in ways chatbots don’t: errors compound across steps, they can loop forever burning API credits, and they’re vulnerable to prompt injection — malicious instructions hidden in a web page or email that hijack the agent. Use guardrails, least-privilege permissions, spending limits, and human approval for anything that sends money, deletes data or emails customers.

Examples

Claude Code and Cursor’s agent mode (coding), ChatGPT’s agent features (browsing and tasks), and no-code agent builders like n8n, Lindy, Relevance AI and Gumloop. For the broader trend, see agentic AI.