Build With Nalo Seed: Setting Up Your Own Free AI Agent

On Thursday, July 2, we ran the latest Build with Nalo Seed session, our monthly hands on hour where we set something up live and you follow along on your own laptop. This month the goal was simple: get a working AI agent running from scratch, for free, and then talk about how to give it real jobs.

If you have been hearing the word "agent" everywhere and are not sure what it actually means in practice, this is the recap for you.

Start Free, Then Grow Into It

The whole point of the session was that you do not need a budget to begin. We downloaded a free desktop agent app, created a free account, and started with a free model so nobody had to put in a credit card to follow along. Once that was working, we added our own API keys to unlock stronger models for the tasks that need them.

That is the pattern we recommend for almost everyone: prove the workflow on free tools first, then spend money only where it clearly pays off.

An Agent Is Not a Chatbot

The first thing we cleared up is the difference between a chatbot and an agent. A chatbot talks. An agent takes action. It can read your email, look something up, update a record, or draft a reply and put it in front of you.

We walked through the same mental model we use with every client, the five layers of an agent: the model (the brain), the harness (the app that runs it), the tools (the hands that do things), the skills (the playbook it follows), and the connectors (the plugs into your real systems like Gmail or your calendar). If you want the longer version of that breakdown, we wrote it up here: The Anatomy of an AI Agent.

Pick the Right Model for the Job

You do not have to marry one AI. Using a service like OpenRouter, you get a single key that lets you reach many different models, Claude, ChatGPT, Gemini, and others, and swap between them depending on the task.

That matters because each model has its own personality. Some are warm and thorough, some are fast and direct. A fun trick we covered: give the same task to two models and let them compete, then keep the answer you like better. During testing you can lean on free models to keep your costs near zero while you figure out what works.

Organize Agents Like a Small Team

Once you have one agent working, the natural next step is to run a few of them like a little company. You have a lead agent that delegates, and specialists underneath it. A research agent, a writing agent, a design agent, each pointed at the tools that fit its job.

You are not building an org chart for fun. You are matching the right model and the right tools to the right task, so each piece of work is handled by the part of your setup that is best at it.

Keep a Human in the Loop

This is the rule we repeat the most, so we said it again. Anything that goes out with your name on it, a client email, a website change, a public post, should get a human look before it ships. The agent drafts, you approve.

And when an agent gets something wrong, the fix is not to give up on it. The fix is to update its instructions so it does not make that mistake again. That is how these systems actually improve over time, you teach them.

Give It the Repetitive Work

The best first jobs for an agent are the boring, repeatable ones. We talked through the idea of a "loop," where an agent quietly watches something like your inbox or a Slack channel, notices when a task shows up, and kicks off the right workflow.

The example we come back to is a morning operations check. Before the day starts, an agent reviews your email, your calendar, your payments, and your open work, then hands you a single organized list of what needs attention and what is still owed. It is the kind of triage that eats an hour of your morning, done before you sit down.

We also showed how you can talk to your agents by voice from your phone, queue up instructions while you are away from your desk, and let them run in the background until you are back.

Then We Built One Live

To make it real, we set up a fresh agent during the session, gave it a calm and precise personality, connected it to Google, and used voice commands to have it scan Gmail and start sorting messages into labels. Watching an agent you built five minutes ago actually clean up your inbox is the moment it clicks for most people.

Protecting Your Data

A big part of the conversation, and a great question from the room, was about keeping sensitive information safe. If your inbox holds client records or HR data, you do not want to point an AI at all of it.

A few things we recommend:

• Give the agent its own email account. Set up a separate, dedicated address for AI work so your sensitive client mail stays isolated from the agent's day to day tasks.
• Scope its access. Point the AI at specific project folders rather than your entire drive, so it only ever sees what it needs for the job in front of it.
• Watch where the data lives. Choosing the "do not train on my data" setting is good, but the data can still sit on a provider's servers. For the most sensitive work, running a model locally on your own machine keeps everything in house. The hardware for that is still steep today, but smaller, capable local models are coming fast, and we expect that to get much more practical over the next few months.

What's Next

Build with Nalo Seed runs every month, and each session takes on a different piece of putting AI to work in a real business. Bring a laptop and you will leave with something that actually runs, not just notes.

If you missed this one, or you know someone who keeps meaning to get started and never does, send them our way. Register for the next session.

And if you would rather have us look at your specific setup and help you decide where an agent fits, book a free consultation. However you want to grow, we would love to help.

Paul