The 3 levels of AI: chatting, agent, agentic (the Claude climb)

Hey, it's QingSu 🌱 All three levels of AI, what each looks like day to day, and the exact setup to reach levels two and three.

the ladder

The 3 levels, in one glance

• 🗨️ Level 1, chat. You open Claude, ask, it answers, then it waits. Brilliant for thinking, useless for doing. A useful starting point.

• 🔌 Level 2, agent. You connect Claude to your real tools (Notion, Gmail, Drive, your calendar) so it can read and act on your stuff, not just talk about it.

• 🤖 Level 3, agentic. You hand Claude a goal, and it plans the steps, does the work, checks its own output, and keeps going until the job is done. No babysitting.

The jump between each level is the difference between saving minutes and saving your whole week. Here's exactly how to climb.

where most people live

Level 1: chatting

Level one is the large language model on its own: Claude, ChatGPT, Gemini. It's trained on a huge amount of text, which is why it's so good at writing, summarizing, and helping you think through a problem. But it's reactive. It only responds to what you type, and it can't touch anything outside the chat window.

What it's great at

• ✍️ Drafting, rewriting, and summarizing.

• 🧠 Thinking through a decision or plan with you.

• ❓ Answering questions and explaining things.

The ceiling: you're still the one doing every click. Claude can write the email, but you copy, paste, and send it. It can plan your week, but you type it into your calendar. That copy-paste tax is exactly what levels two and three remove.

plug it in

Level 2: agents (connect Claude to your tools)

An agent is that same model connected to real tools through a connector, built on MCP (Model Context Protocol), the open standard Anthropic created for AI apps to talk to your data. Once connected, Claude can read data from your apps, and create, modify, or delete things on your behalf. Source: Anthropic connectors and MCP help.

How to turn it on (5 minutes)

1. 🔧 In the Claude app or on claude.ai, go to Settings, then Connectors (Customize, then Connectors on Pro and Max).

2. ➕ Pick a ready-made connector from the directory (Notion, Google Workspace, Slack, and more) and click Connect, or click Add custom connector to paste your own MCP server URL.

3. 🔑 Sign in to the tool and grant Claude access to only the pages or accounts you choose.

4. 💬 Back in a normal chat, just ask Claude to use it. For this exercise, start the task yourself in chat and review the actions Claude proposes.

Custom remote MCP connectors are available on Free and paid plans; Free is limited to one custom connector. Availability and permissions vary by connector and workspace. Source: Anthropic connectors and MCP help.

What each connector can do

📄 Notion read AND update pages and databases, so Claude can pull a doc into the chat as context, then edit it or add a new row on your say-so. Source: Notion MCP help.

📧 Gmail search, read, draft, send, reply, and forward. Sending actions ask for approval by default; Team and Enterprise policies can allow different approval settings. Source: Claude Google Workspace connectors.

📁 Google Drive and Docs Drive retrieves files, creates folders, and saves generated files when code execution and file creation are enabled. Embedded document images are not processed during text extraction. Separate Docs, Sheets, and Slides connectors support live editing. Source: Claude Google Workspace connectors.

📅 Google Calendar view, create, update, and delete events, and find mutual free time. Source: Claude Google Workspace connectors.

Try this prompt once you're connected

🔌 Your first connected task

Using my Notion connector, read my content calendar database, find every post still marked "idea", and turn the three strongest into filled-out rows with a hook, a caption, and a suggested post date. Show me the rows before you save anything.

Heads-up: connecting a tool and scheduling a task are separate setup steps. This example starts in chat; scheduled workflows need their own configuration. Review each connector's available actions and approval settings before letting it change anything.

hand over the job

Level 3: agentic (hand over the whole job)

Level three is where Claude stops waiting for your next instruction. You give it a goal, and it plans the steps, does the work, checks its own output for errors, and iterates until the task actually meets the goal. It handles multi-step work start to finish without you standing over it.

The clearest place to see this is Claude Code, Anthropic's agentic tool that reads your files, edits them, runs commands, and works through multi-step tasks autonomously. Describe what you want in plain language, and it plans the approach, writes the work across multiple files, and verifies it works, then fixes what broke. Source: Claude Code overview.

What makes it agentic, not just automated

🧭 It plans. Plan mode lets Claude map out the steps before touching anything, so you approve the approach first.

🔁 It self-checks. It runs its own work, spots errors or failed tests, and loops back to fix them before reporting done.

👥 It can split up. You can delegate focused parts of a job to subagents, each with its own context, and use their results in the main session. Source: Claude Code subagents.

⏰ It can run on a schedule. Cloud scheduled tasks (formerly routines) run on Anthropic's infrastructure so tasks like a morning review keep running even when your computer is off. Source: Claude Code overview.

Getting started with Claude Code

Claude Code runs in your terminal, VS Code, the desktop app, or your browser at claude.ai/code. On Mac, Linux, or WSL, the official installer command is:

💻 The one-line install

curl -fsSL https://claude.ai/install.sh | bash

Then run claude inside any project folder. Sign in with an eligible Claude subscription or an Anthropic Console account; the CLI also supports selected third-party providers. Account requirements vary by surface. Source: Claude Code overview.

You don't need to be a developer to feel it. Point it at a folder of messy files, a spreadsheet, or a repetitive weekly task, give it the goal, and watch it work the list while you go make a coffee.

go further

3 bonus prompts to run next

1 · Spot the copy-paste tax

Find your level-two win: Look at everything I did on my computer this week. List the five tasks where I was just moving information between apps (copy here, paste there), and tell me which Claude connector would remove each one.

2 · Scope your first handoff

Design your first agentic task: I want to hand off one repetitive weekly job completely. Interview me with five questions to understand it, then write the exact goal and step list I'd give an agent so it can run the whole thing and check its own work.

3 · Choose your connect order

Pick the right connectors: Here's my tool stack: [list your apps]. Tell me which ones Claude has a connector for today, what each connector can and can't do, and the order I should connect them in for the biggest time saved.

use it well

How to get the most out of it

🌱 If you're just starting don't try to jump straight to level three. Connect one tool you use every day (Notion or Gmail is a soft landing) and let Claude read and draft for a week before you trust it to act. The confidence compounds fast.

⚡ If you're already comfortable the real leverage is chaining levels. Let an agentic session read your connected tools (level two) as part of a bigger goal (level three), so it gathers its own context instead of you feeding it. That's when one prompt replaces an afternoon.

match the stakes

An honest note

More autonomy is not always better. For anything high-stakes or irreversible (sending money, emailing a client, deleting records) keep Claude at level one or two and stay in the loop. Agentic shines on repetitive, low-risk, multi-step work where a small mistake is cheap to catch. Match the level to the stakes, not to the hype.