Guides › New to AI

AI agents, explained for normal people

Updated Oct 22, 2026. No hype.

You've probably seen posts claiming that "99% of people use AI like Google" while a clever 1% run swarms of 65 "self-learning" agents. Here's what an AI agent actually is, what "multi-agent" and "self-learning" really mean, and what the companies that make these tools say about how many you need.

An AI chat vs an AI agent

A chat is a conversation. You ask, it answers, you ask again. Nothing happens unless you type.

An agent is the same AI given a goal, some tools and permission to take steps. It can search the web, click through websites, use connected apps like your email or Google Drive, create documents and spreadsheets, and keep working in the background. Some can run on a schedule, like a briefing every morning.

Agent work isn't free extra. Anthropic, for example, says Claude's usage limits are shared across every Claude app and Claude Code, and multi-step tasks use more of them. ^5,6

What "multi-agent" really means

What does multi-agent mean?

An AI can only keep so much in its head at once (its context window). Long jobs, like reading 40 web pages, fill it with clutter. The fix is to hand that side job to a helper agent (Anthropic calls them subagents). The helper works in its own fresh workspace and hands back only a short summary. ³

• Helpers aren't free. Each subagent counts toward the same usage limits as the main conversation. ³
• There are limits on purpose. Claude Code runs at most 20 subagents at the same time by default. ³
• "Agent teams" are a separate, experimental feature. In Claude Code they're switched off by default. Several AI sessions share a task list and message each other. Anthropic says they add coordination overhead and use significantly more tokens than a single session, about 7 times more when teammates work in plan mode. ^4,7

The everyday version of this, outside programming, is a chief-of-staff assistant: one AI you talk to, which passes jobs to a few specialists (research, writing, bookkeeping) and brings the results back to you. How to set that up without coding.

What "self-learning" really means

What does self-learning mean?

No consumer AI agent rewrites its own brain from your work. When a tool "gets smarter over time", it's one of these four things:

1. Memory. The app saves short notes about you (your role, preferences, ongoing projects) and reads them in future chats. Claude's memory is on by default for Free, Pro and Max, and each Claude Project keeps its own memory. ¹
2. Instruction files. A written list of house rules the AI reads at the start of every session: project instructions in the Claude app, or a CLAUDE.md file in Claude Code. Claude Code also has "auto memory", notes Claude writes itself when you correct it. ^8,2
3. Skills. Saved step-by-step how-tos the agent loads only when it needs them, like "how we format the monthly report". ⁹
4. Routines. Scheduled tasks that repeat a job on a timetable, like a weekly price check. ^10,11,12

That's genuinely useful. But it's note-taking and checklists, and notes can go stale or be wrong. Anthropic's advice for Claude Code is to keep instruction files short (under about 200 lines) and delete anything that's no longer true. ^2,13

The honest truth about "65 agents" claims

The post that sparked this guide claimed most people "use Opus 5.5 like Google" while the top 1% run "swarms of self-learning Claude agents" with chief and project-manager agents. The video it shared is a real Anthropic workshop, Maximizing value from Claude Code²¹, presented by Lydia Hallie of Anthropic. In that video, Anthropic doesn't mention 65 agents, self-learning agents or chief agents. It covers much more ordinary, useful things:

• What gets sent with every request (your instruction file, skills, connected tools), and why that costs tokens
• The levers that control cost: how much you send, how much comes back, the effort setting, and which model you pick
• When to try a stronger model and when to raise effort instead
• Clearing or summarizing long conversations, and why a break can make the next message cost more
• Handing noisy side jobs to subagents
• Usage limits that team admins can set

We turned those into a plain-English guide: Get more out of your Claude plan for less.

What Anthropic's own docs say about agent numbers:

• "Start with 3-5 teammates for most workflows." Even with 15 independent tasks, Anthropic suggests 3 teammates as a starting point. ⁴
• More teammates means more coordination and more chances for conflicts, and beyond a point extra teammates don't speed things up. ⁴
• Agent teams use significantly more tokens than a single session, and the advice is to keep teams small, use the cheaper Sonnet model for teammates and shut them down when they're done. ^4,7
• Claude Code caps running subagents at 20 at a time by default. ³

Our take: the number of agents tells you nothing about the results. A setup of one assistant plus a handful of specialists, each with a clear job and a human approving anything that leaves the building, beats 65 agents you can't afford or check.

What you can try today

Most paid plans already include some agent features. Prices are from our plan data (official prices only, USD and CAD never converted).

Full comparison with sources: AI agents compared.

Guides in this series

• Get more out of your Claude plan for less: fresh chats, model vs effort, Projects, and the Claude Code commands that save the most
• How to run a small team of AI agents (without being a programmer): a chief-of-staff assistant, specialists, routines and approval gates
• What Grok Bot can actually do for you: powerful without the hype, with honest limits

Common questions

5 quick questions, answered

What is the difference between an AI chat and an AI agent?

A chat answers you, one message at a time, while you watch. An agent takes a goal and works through the steps itself: searching, opening websites, using connected apps, making files, and sometimes running on a schedule while you're away. It still needs you to check the result and approve anything that gets sent, bought or deleted.

Do self-learning AI agents really teach themselves?

Not in the way the hype suggests. The AI model doesn't retrain itself on your work. What these tools call learning is memory: notes the app saves about your preferences, instruction files you or the agent write, saved how-tos called skills, and scheduled routines. That's useful, but it's note-taking, not a brain upgrade.

Do I need 65 AI agents?

No. Anthropic's own Claude Code docs suggest starting with 3 to 5 teammates for most work, cap running helpers at 20 at a time by default, and warn that agent teams use significantly more tokens than a single session. More agents means more cost, more coordination and more to check.

Do AI agents cost extra?

Usually not a separate fee, but they use more of your plan. Agent work counts toward the same usage limits as chatting, and long multi-step tasks use up limits faster. Some agent features are only on paid plans.

Which AI agent should a beginner try first?

Start with whatever your current plan already includes: scheduled tasks in Claude or ChatGPT, or scheduled actions in Gemini. Give it one small, repeating job, like a morning summary, and check its work for a couple of weeks before adding more.

Related: AI agents compared · Get the most out of AI · Which AI model should I use? · AI crash course

New to AI?

AI crash course: start here

What AI chatbots are, how to ask, and how to stay safe. About 10 minutes.

Should I pay for AI?

What free plans really give you, what a paid plan adds, and who needs one.

Which AI model should I use?

Pick the right model and thinking level in 60 seconds.

Tip: subscribe on the website, not in the app

Buying through the iPhone or iPad app often costs more because of Apple's App Store fees. Sign up on the provider's own site (like claude.ai or chatgpt.com) and you can still use the app with the same account. How to switch and what to know · Compare app and website prices.