Build Your First AI Agent in One Hour

An AI agent at the non-technical definition, think of it as a chat bot that you give instructions as well as access to certain systems, perhaps calendar, email, Slack, Salesforce, et cetera.

So once an agent has access to instructions as well as systems, then they're able to complete the workflow for you.

It can advance to complete a task for us.

So what that means is that if it's going to be able to decide, Hey, I need to check calendar. I need to check email and create an agenda for this meeting or send a follow-up recap, then it needs to have access to systems.

I have a non-technical background.

Over the summer, I started teaching a group of friends to build agents in one hour. Since then, over 300 people have joined me, and I am most excited to be with WAAI today for us to build our agents.

One of the great things about a system like Gumloop is you can move your model for your agent from one large language model to the next in one click.

And if one model becomes too expensive, you can actually slide over to the next model easily.

If you have a routine that you repeat frequently, you can customize this agent to take on that repetitive task.

But in general, we're going to start off with a chief of staff agent today. That's something most people can relate to. But if you came into this session saying, Jacqueline, I know exactly what I'd use an AI agent for, feel free to deviate slightly and you'll see where you can make those deviations to get your agent up and going.

There are three rules that I like to follow as a best practice for setting up an AI agent.

Number one, give the agent a persona. So for those of you who are building a chief of staff agent alongside us today, we're going to first tell our agent, you are my chief of staff agent. So we give our agents a persona. The next thing that we want to do is we want to tell the agent what its goals are. And this is for any agent, persona, goals, and deliverables.

I tend to find that speaking into the microphone for one to two minutes gives my system enough information to start working.

We can type it all out or we can speak into it. What you now see on my screen is that the microphone actually translated into text. I probably would have only typed a sentence or two, especially with the fancy nails that I have on.

What you'll see on your screen is that your agent starts to think.

So what's going to happen next is our screen is going to change slightly. Do not be alarmed. This is a regular part of the build process. And you'll see that it actually dictates and transcribes what it's doing based on the instructions that you provided on my screen. If you're following along on my screen, my agent is now asking for input. It's provided multiple choice questions for me. I can select from those options, but if none of those fit, I can use option D, which is other, and I can type in my response. So if you have questions on your screen, I'd like you to go ahead and answer those questions and hit submit. Your agent will begin to think again.

If you're looking at my screen, I've given my agent a name and this one has nothing to do with function.

It's all about aesthetics. When I log into my dashboard, I come in to the icon section and select an emoji with color. Again, this has nothing to do with function. It just gives me a little bit of flair when I come onto the screen.

I would suggest if you're unsure safe space to start is one of the Google models.

I typically build my new agents in Moonshot using the Kimi K3 model. It's one of the Chinese large language labs, Kimi K3 is one of my favorite in terms of setting up agents. The reason why I like the Kimi K3 model. It tends to give me great results and it doesn't use a lot of tokens inside of Gumloop. If you're unsure about which lab to go with, Google's doing a pretty good job right now, I'd say that's a safe starting point.

What that does is it allows the agent to learn from your interaction inside of the chat feature.

For example, if you click new task that opens another chat inside of this agent, what this blue button does is it allows your agent to learn how you're interacting with it and it automates updates.

As a best practice, I have a 15 minute meeting that reoccurs every three weeks on my calendar where I jump in and I review my agent instructions so that as it's auto updating, I am still fully aware of what's happening.

If you're using your agent daily, you may want to check yours every two weeks. Most of my agents, I have them running a couple of times a week. So a three week cadence works for me. But I just want to make you aware that if you keep the edit on own instructions on, you want to be able to manage and have oversight into what the agent is updating. So please put a calendar reminder at some frequency for you to come back and check that.

Connectors are what make agents work.

You'll see something called connectors.

Exa is similar to Google, but it's built for agents.

So it means that as agents are searching the web and fetching information, searching the internet and reading web pages that search and fetch, it is built natively for agents to be able to pull the information that humans are looking for.

Apify is an agent native tool that fetches specific information from various platforms such as LinkedIn.

Now Apify looks at other websites, but I found it to be one of the best if you're prepping and you're using LinkedIn to prepare for a meeting, Apify is one of the best tools for that because they have written a protocol for agents to be able to look at public pages.

Firecrawl, similar to Exa and Apify, it is a system built to help agents to be able to process information on the web.

So similar to how I would go to Google and search for information, my agent would first go to Firecrawl and to be able to go through websites and pull information on my behalf. If I'm meeting with an organization, have they been in the press recently? Have they shared a press release? Is there anything they're celebrating? I use my agent's superpower tools, these connectors, to help me prep for that.

LinkedIn doesn't share the MCP and API.

I'm getting a little technical here, but that's simply how one system talks to another. LinkedIn doesn't readily share that information. Now Apify looks at other websites, but I found it to be one of the best if you're prepping and you're using LinkedIn to prepare for a meeting, Apify is one of the best tools for that because they have written a protocol for agents to be able to look at public pages.

Skills are actually workflows. It tells our agent what to do step by step.

We can create skills manually, meaning we can tell as if we were writing SOP for a new hire, we could write that for the agent, we could also use something like a site, what I'm showing on my screen, skills.sh, or we can have the agent create a skill for us.

I am very hesitant to copy someone's skill offline if I don't have access to read it before I download it.

Prompts can be a few sentences or a few paragraphs. Skills can be pages. People can write any set of instructions in a skill and you can inadvertently give away access to data that you don't feel comfortable with.

If I notice that a skill has been installed 3.7 million times, I may poke around to understand versus a skill that's not been downloaded at all, I won't even poke around and look at that.

The way that I use this site, skills.sh, I kind of think of it like I'm going to a restaurant and I'm looking for reviews. I may not look at the reviews if there are only five reviews. I may not look at the restaurant if it has 3.6 stars. However, if I see that a restaurant has 500 reviews and 4.9 stars, I then dig in deeper and read a couple of the reviews before I decide I'm going to book a reservation.

The best way to create skills inside of Gumloop is to come back to the dictate feature.

What I typically do is I start off by saying I want to create two skills for the COS agent. Ask me three or four questions to customize my agent's skills. It's just that simple.

The next thing that I'm going to do is I'm going to tell my agent to update agent instructions.

So if you're paying attention to my screen, under the Kimi K3 model that I selected, there's no instructions for the agent yet. So just with three words, update agent instructions, it's going to start thinking again.

One of the things that I like to do with my agents is I give them guardrails. I tell them what they can do and what they can't do.

And this is just as important of telling your agent what you want it to achieve.

One of the cool things about this platform is compared to other AI systems, you get a lot of control over what it has access to.

So under tool management, I'm going to select read only tools. There are six options underneath there. I can then restrict or give permission to my agent on what I wanted to read only. Beneath that section, there are write/delete tools. And in that section, I can then give permission or restrict access to sending emails, updating, creating drafts, forwarding emails, et cetera. One of the reasons why I like this feature, many of the other tools don't yet provide this level of detail when you're giving your AI system access to a tool on what you can restrict and give permissions for.

If I click on the three dots, it allows me to switch an account, but it also gives me access to easily remove access.

On the connectors section under Gmail, on the far right side, there is a circle with a check mark and next to it three dots. So there, I can easily remove access from Gmail. So just like that, my Gmail access has been removed.

It's also important for me to note here, if you change your password in any system that you've given your Gumloop access to, your Gumloop will no longer be able to access that system.

So you would have to come back into Gumloop and reauthorize that permission for it to continue to work.

Don't say try again, actually say, "Hey, you surfaced this and this isn't beneficial and this is why."

The next thing to do with our agent is run a test. So another three word prompt. Run a test is where we're allowing our agent to then think and provide us with our first result to which we should give feedback. So if your agent comes back and it gives you a result that you love, tell it why you love the results. If your agent comes back and it gives you a result and you have mixed feelings about it, then you'll tell it, "Hey, I like this, but I don't like that," and tell it why. If you absolutely don't like the results, also tell the agent why. Again, if we think about how we onboard a really competent new hire, the level of instruction we would provide, we'd want to do something similar for our agents.

681 credits is not a lot for setting up agents.

Gumloop uses credits instead of tokens and inside of the Gumloop university section, it explains that it calculates credits similar to how OpenAI and ChatGPT and Gemini calculate tokens.

To delete this, next to the box that has an up arrow and a half open box, there are three dots. If you were to click on the three dots, you would be able to delete the agent.

If you see an error, what I recommend that you do is refresh your screen. Sometimes when you're deleting an agent, it actually already went through and deleted. So the first thing I would do is refresh. In most cases, that actually will confirm that the agent has been deleted. One of the reasons why, another reason why I appreciate the Gumloop team, they're not trying to trap you in or trick you. They make it really easy to remove access and really easy to delete agents.

What a lot of folks are doing is setting up an agent account and only forwarding information that they want the agent to read.

Yes, it is an extra step, but it's a great way to be comfortable with the level of access that you're giving to a system. So you can select a dummy account. What I recommend just starting out, most of us have an email account that we send like promotional emails, spam emails to just use that account to get started.

It depends on the level of permissions we give it.

SOC 2, GDPR and HIPAA, there is compliance there.

One of the things that they put together is their trust center screen. And typically when enterprise level legal teams join my trainings, this is one of the resources that they use as they begin their vetting process to give their organization access to an orchestrator like Gumloop. Gumloop is used by Shopify, Instacart and many other large enterprises.

I don't work for Gumloop.

I found this tool over the summer through Lenny's podcast, and I started using it and have since made myself an ambassador for them. I've created a relationship with Gumloop.

A best practice is go ahead and also put a note in your phone to cancel in 12 days if you're only doing a trial run. That way you don't get a surprise charge in 14 days.

Well, there's two ways to do that. So one, I would actually cancel the subscription and then I would delete the account.

As it's reasoning, you'll see exactly what it's doing.

You can double-click into those later into the reasoning and review that. That's one of the reasons why I like building in this platform. You can build agents inside of ChatGPT and Claude. They don't always allow you to click into the reasoning.

I can now tell the agent I'll pause on that so it won't continue to prompt me to ask for that.

I am actually going to give my agent access to my calendar. You'll notice I didn't connect it when I was in the connector section, but I was able to one-click log in here and I'm not going to connect mine to Slack now, so I'm also going to say I'll pause on connecting Slack. In my initial prompt, I said I'm going to use Slack, but now I'm pausing.