AI Guide

What are AI Agents?

AI Agents (subagents) are “specialized AI helpers” — each with its own role, tools, and job — that run in their own window, and the main AI calls them in automatically when it meets a matching task. In short, they turn a single AI that does everything itself into a “team” that hands specialized work to the right helper.

Updated July 2026·~7 min read·Level: Beginner

What AI Agents really are

An AI Agent (or subagent) is a specialized AI helper set up with a clear role and job — like a “code reviewer” (code-reviewer), a “debugger”, or a “researcher”. Each one runs in its own window (context), with tools limited to just what that job needs, then hands its result back to the main AI.

Think of the main AI as a “team lead” — whatever task comes up, it hands it to the agent best suited for it, then collects the results and answers you, all without cluttering the main conversation.

How AI Agents work

Each agent is a short “definition” file that states 3 things: its name, what it does and when to call it (description), and which tools it may use (tools) — followed by the agent's role instructions:

---
name: code-reviewer
description: Reviews code for bugs, vulnerabilities, and improvements
tools: Read, Grep, Glob
---

You are a strict senior code reviewer
When given code, check for:
1. Bugs and missed edge cases
2. Security vulnerabilities
3. Hard-to-read or duplicated spots

Report as a list, ordered by severity, with fixes

Roughly how it goes in practice:

You set up the agent definition file (a one-time setup)
The main AI sees the list of agents, with descriptions of what each is good for
When you give a task that matches an agent's job, the main AI “delegates” it to that agent
The agent works in its own context with its limited tools, then returns a summary for the main AI to turn into an answer

💡 Key point: An agent has its own separate context, so its long work doesn't clutter the main conversation — and because it only gets the tools it needs, it stays focused on one job and is safer.

What AI Agents let AI do

AI Agents turn a single AI into a team, splitting complex work into sub-jobs each with its own specialist — for example:

Code review — a specialized agent hunts for bugs and vulnerabilities and reports them by severity
Debugging — an agent focused purely on finding the cause of an error
Research / lookup — an agent that sweeps many files or web pages and hands back a short summary
Writing tests / docs — offload repetitive work to a specialized agent
Working in parallel — hand several sub-jobs to several agents at once, then combine the results

AI Agents are specialized helpers that run on their own, while a Prompt sets the main AI's role and AI Skills add specialized know-how it draws on.

Real-world AI Agent examples

Specialized agents people actually use:

code-reviewer — reviews code for bugs, vulnerabilities, and improvements
debugger — tracks down the cause of an error and proposes a fix
researcher — searches and summarizes across many files/sources, concisely
test-writer — writes a test suite for given code
doc-writer — drafts docs or a README from the real code

See curated Agents, with clear notes on what each does and how to set it up, in the Agents library.

Does it work with Claude, Cursor and Codex

Yes — each tool has its own place to set up agents. The “specialized helper” idea is the same; only where you put the file/config differs:

Claude Code — drop the definition file in .claude/agents/ and the main AI calls it as a subagent automatically
Cursor — set it up as a custom mode / specialized role in the editor
ChatGPT — build it as a Custom GPT split by job

Same idea across tools — write the agent's role and job once, then adapt where it's configured to fit the tool you use.

How to start using AI Agents

Setting up your first agent takes 3 steps:

Pick an agent that fits the job — want a code reviewer? Choose code-reviewer. Open the Agents library and browse by task.
Place the definition file in your AI tool — copy the agent file (name/description/tools + role) into each tool's location, e.g. .claude/agents/.
Try a task that matches its job — give a task that fits the agent; if set up right, the main AI calls the agent on its own.

Don't want to set them up one by one? Pick a Loadout with Agents — bundled by task, ready to use.

Are AI Agents safe

Agents are fairly safe by nature, because they only get the tools they need and run in a context separate from the main work. Still, keep these in mind:

Limit tools to just what the job needs — a code-review agent only needs Read/Grep; no need to grant file-deletion or message-sending rights
Use agents from trustworthy sources — always read the role and the tools list in the file before setting it up
Watch tasks with real impact — agents that can edit files or run commands; check their output before relying on it

✅ We curate only well-written agent definitions with clear roles and tools, plus notes on which suits which job, so you can choose with confidence.

AI Agents vs Prompt vs Skill

These three are often confused, but they do different jobs:

Difference	AI Agents	Prompt	Skill
What it changes	Adds a specialist helper that works on its own	Sets the main AI's role/instructions	Adds specialized know-how
Where it runs	In its own separate context (window)	In the same main chat/context	In the main context, when the AI draws on it
Who triggers it	The main AI calls it automatically on a matching task	You paste the prompt yourself	The AI picks it up itself on a matching task

Read on: What is a Prompt · What are AI Skills · What are AI Rules

Loadouts with AI Agents

In practice you rarely use one agent alone — you combine several with MCP/Skills/Rules/Prompts to finish one job. A bundle assembled like that is called a Loadout.

For example, a web-dev Loadout might include a code-reviewer + debugger agent, MCP for GitHub, and a coding-standards Rule in one set — ready to use without assembling it yourself.

Next steps

Browse the Agents library Pick a Loadout with Agents How to install Agents Agents recommended by task Build your own Loadout