Give your agent tools

Complete Your first agent: give it one job first. You need a working agent Project to test tools inside.

Without tools, your agent is stuck inside a text box. With tools, it can read your email, check your calendar, search the web, and update your files. The agent you built in the previous lesson works, but it only knows what you paste into it.

Tools are how you give it eyes and hands so it can go get information and take action on its own.

A tool, in agent terms, is a specific capability you plug into the agent. "Read my Gmail inbox" is a tool. "Search the web" is a tool.

Each tool does one thing, and the agent decides when to use it based on the job you gave it. You do not tell the agent "now use the email tool." You describe the job ("triage my morning inbox"), and the agent figures out which tools it needs.

Here is how the decision works inside the agent's loop:

The agent reads your request (perceive). It looks at the tools available to it, like a worker checking which equipment is in the shed, and picks the one that matches the next step (decide). It calls the tool, gets a result, and uses that result to continue or finish (act).

If the email tool returns 14 unread messages, the agent reads them, sorts them, and then maybe reaches for the calendar tool to check if any meeting conflicts need flagging.

The key idea is that the agent chooses the tool, not you. You describe the outcome, and the agent picks the path. This is what separates a tool-using agent from a chatbot where you manually copy-paste between apps.

Three real tools available right now (no code, no terminal):

1. Web search. Claude (free and paid), Antigravity, and most AI platforms can search the web during a conversation. The agent perceives your question, decides it needs current information, and searches without you pasting a URL. 2. File reading. The Claude desktop app (Cowork) reads files from your disk. The agent perceives the file contents, decides what to extract, and produces a summary or analysis. 3. Connectors (MCP). MCP stands for Model Context Protocol, the standard way to plug external services (Gmail, Google Calendar, Notion, Slack, and many more) into an agent. Connectors let the agent read from and write to your real tools. The hands-on wiring lesson for MCP is Wire Your First MCP in the core path. This lesson teaches the concept; that lesson teaches the install.

Why this matters more than it sounds: each tool you add multiplies what the agent can do. One tool (web search) makes it a researcher; two (add file reading) make it an analyst; three (add a connector to your CRM) make it an operator.

The tools are the multiplier. More tools, more reach, more work the agent handles without you in the middle.

The tools themselves are not magic. They are specific, named capabilities with clear limits.

The web search tool cannot read your private email. The file reader cannot send messages.

Each tool does one thing well, and the agent's job is to combine them in the right order. When a tool fails or returns something unexpected, a well-instructed agent should tell you what happened instead of guessing.

Pick the agent Project you built in the previous lesson. Open a new conversation inside it and try this:

STEP 1: Test web search (built in to most AI surfaces)
Ask your agent a question that requires current information it could not know from your instructions alone. For example: "What is the current price of [a stock or product you track]?" or "What happened in [your industry] this week?"

If the agent searches the web and returns a result, it used a tool. **Watch for the moment the agent reaches for web search on its own**, without you telling it to. That is the decide-to-call pattern in action. If it says "I don't have access to current information," web search may be toggled off in settings. Toggle it on and try again.

STEP 2: Test file reading (desktop app or file upload)
If you are using the desktop app, point your agent at a file on your disk (a report, a spreadsheet export, meeting notes). If you are using claude.ai, upload a file to the conversation. Ask the agent to summarize it using the perceive-decide-act structure from your instructions.

STEP 3: Map your ideal toolkit
Ask Claude: "Based on the job this Project does, what three tools would make me most productive? For each tool, describe what it would perceive, what decisions it would help with, and what actions it would take. Be specific to my use case."

This gives you a personal tool roadmap. You do not need to install anything yet. The map is the takeaway.

Based on your Weekly Email Digest agent, here are three tools that would make you most productive:

1. **Gmail connector**
   - Perceive: reads your inbox directly (no more copy-pasting emails)
   - Decide: applies your urgency categories automatically on fresh messages
   - Act: drafts replies in your voice and queues them as drafts in Gmail

2. **Google Calendar connector**
   - Perceive: checks your schedule for the next 48 hours
   - Decide: flags emails that conflict with calendar events (meeting reschedules, deadline changes)
   - Act: suggests calendar updates alongside the email triage

3. **Notion or Google Sheets connector**
   - Perceive: reads your project tracker or task board
   - Decide: cross-references email topics with open tasks
   - Act: adds new action items from emails directly to your tracker

Each connector plugs in through MCP. The wiring lesson is at /lessons/l7-01 when you are ready to install them.

What good looks like

• You tested at least one built-in tool (web search or file reading) inside your agent Project
• You observed the agent choosing to use the tool on its own, not because you told it to
• You have a three-tool roadmap tailored to your specific agent's job
• You can explain what a tool is in agent terms (a specific capability the agent decides when to use)

Open Claude (outside your Project, in a fresh conversation) and paste this:

"I want to understand how an AI agent decides when to use a tool. Give me a scenario where an agent has three tools available (web search, file reader, and email sender) and receives this request: 'Prepare a weekly market update for my team based on the latest news and our internal sales report, then email the summary to the team list.'

Walk through the agent's decision process step by step. For each step, name which tool it picks and why. Show me what happens if one tool fails (the email sender is down) and how the agent should handle it."

This exercise builds your intuition for tool selection and failure handling, which matters once you start connecting real tools.

What good looks like

• Claude produced a step-by-step walkthrough of the agent choosing tools for a multi-step job
• The walkthrough includes at least one tool-failure scenario and a recovery path
• You can explain in your own words how the agent picks which tool to use at each step