Build an AI Agent · Lesson 2Pro~14 min readTools + multi-step workflowsBuilds on Lesson 1

Give your agent tools — and let it act, not just talk.

A talking agent can plan; a connected one can do. This lesson is about giving your agent hands — search, files, email, spreadsheets, your apps — and chaining them into a workflow that actually moves work forward, all while keeping you safely in control.

The mental model

An agent is only as capable as the tools it can reach. Tools are its hands.

A talking agent can plan and summarize, but it can’t change anything in your world. Give it tools — search, your files, email, a calendar, a spreadsheet, an API — and it goes from advising you to acting for you. The skill is connecting the right tools, safely, and chaining steps into a real workflow.

The Reframe

Start read-only. The fastest way to trust an agent is to let it look before it touches. Give it read access first; add the ability to write, send, or change only once it’s proven on the read-only version.

Step 01 Tools are the agent’s hands

Common tools worth connecting, roughly in order of safety:

Search & the web — research, fact-finding, monitoring.
Your files & docs — read and summarize what you already have.
Email & calendar — read for triage; later, draft and schedule.
A spreadsheet or database — read records; later, add rows.
An API or app — connect to a system you already use.

Step 02 Connect tools safely

Every connection is a decision about trust. Two rules keep you out of trouble:

Read before write. Grant read-only access first; confirm the agent behaves; only then allow it to change things.
Least access. Connect only the specific inbox, folder, or sheet it needs — not your entire account.

Step 03 Chain a real workflow

Real value comes from sequencing tools: gather → process → produce → place the result. Describe the chain explicitly.

Workflow promptRun this workflow: 1) Read [source] and pull [the data]. 2) For each item, [process — summarize / categorize / score]. 3) Produce [the output — a table / draft / report]. 4) Put the result in [destination], but show it to me first. If a step fails or returns nothing, stop and tell me rather than guessing.

Step 04 Patterns for reliable tool use

Name the tool for the job — “use search for current facts, use the file for our internal numbers.”
Say what to do on failure — stop and report, don’t improvise.
Ask before leaving the building — anything that sends, posts, pays, or deletes gets a confirmation.

The expensive mistake is granting send/write/delete access early. Keep a human in the loop for anything that leaves your control — emails sent, money moved, records deleted. Reading is safe; acting on the outside world is where you stay in the loop.

Your challenge: build a connected workflow

Build an agent that does a real chain of work:

Connect one source the agent can read (a folder, sheet, or inbox), read-only.
Define a gather → process → produce → place workflow.
Have it show the result before placing it anywhere.
Run it on real data and check each step held up.

That’s an agent doing genuine multi-step work with your tools. Next, make it reliable enough to trust unattended — guardrails, testing, and oversight. That’s Lesson 3.

Next: reliable agents →Back to Build an AI Agent

What you can do now

Identify which tools an agent needs and in what order of safety
Connect tools with read-before-write and least-access rules
Chain a gather → process → produce → place workflow
Write prompts that pick the right tool and fail safely
Keep a human in the loop for actions that leave your control

Pro+

Up next in Build an AI Agent

Lesson 3 · Reliable agents — guardrails, testing & running unattended

The difference between a demo and an agent you trust: guardrails, real testing, oversight, and earning the right to run on a schedule. Go to Lesson 3 →