A few days ago, I posted this video where Google was announcing the AI agents that they will be incorporating into all of their systems to help us with customer service, to Help us with shopping and to automate our various online tasks that we do day to day.

And today this thing got released into the wild for all us to try out and see how well it works.

They're calling it the agent builder.

I haven't tried it.

I'm going to do it live with you today.

I'm going to go through it, try to Figure out and show you how well it works or doesn't work.

I have no idea how good it is, but it is exciting because Google is promising to have released a no code AI agent builder, meaning that anyone regardless of their tech Background will be able to use it to build their very own agent for their business or for their personal productivity.

With that said, let's dive in.

To try it out, first of all, find Vertex AI at cloud.google.com/Vertex -AI that takes us to this page.

We're going to try it in console that takes you here and I'm going to search for agent, but here's agent builder.

Can we spy on your stuff?

Sure.

Why not?

Then it takes a few seconds to set up the environment.

We have a few choices here.

We have search chat recommendations as an create content recommendation engine, but this agent built using natural language agents can answer questions from data connected With business systems through tools and more and more create them agent.

We're going to call it agent zero.

What else?

Then we click create new app.

It crashed a bunch of times or it just kind of didn't progress.

I'm trying different regions, different names.

Hooray.

So that took, that took some time.

It crashed a few times.

I might've accidentally spun up a bunch of these, but that's okay.

And by default the generative model is Gemini 1.0 pro, but we can change the model used by the published by the completed published agent in settings.

I just saw this.

Cloud three Opus is in public preview on Google clouds, Vertex AI.

Hopefully we'll Figure out how to add it here.

We're going to make this into a, and this is going to be a sales assistant and the goal is going to be help customers order from my website instructions.

You're in instructions.

We're going to say you're a helpful assistant that helps customers purchase items from my online store.

And here we're able to use tools to help this agent, as well as calling other agents to help this agent.

I'm going to go ahead and copy and paste it in here.

We've added, greet the users, then ask how you can help them today.

Summarize the user's request and ask them to confirm that you understood correctly.

If necessary, seek clarifying details.

Use tool to help the user with their task.

Use agent to help the user with a complex subtask.

Thank the user for their business and say goodbye.

In order to use this functionality, we need to define those things first to the tools and the agent.

I'm going to go ahead and delete that.

Oh, and this is save.

I have to click save before I can move around.

With tools, we can use open API data store function.

It looks like to start using tools, you will have to know a little bit about APIs and how to connect different functionality.

This might be more oriented at web developers, which is unfortunate.

Hopefully they can make this easier and more accessible to people, but they do have integration.

Sort of some pre-built applications here mostly for messaging.

I spent about an hour messing around with this stuff and I was able to build a few agents and get them to connect and to actually carry out certain tasks where they're able To pull in other agents as needed.

As an example, here's a sales assistant agent.

And then there's a code agent.

A sales assistant agent, we're telling them you're a helpful assistant that helps customers purchase items from my online store and to calculate how much their order costs And then use this agent code.

That's the second agent that I've built.

We're calling that agent to solve some of the tasks for us that we wanted to solve, right?

We're calling the code agent.

If customers ask how much something is to calculate sales tax and totals, and then I'm telling it, do not do the calculations yourself.

And then if we go over to the code agent, the code agent, their task is help calculate taxes and totals, right?

Like that robot from Rick and Morty, right?

This one passes the butter.

Instructions, another agent will call on you to ask how much the total is.

You know, I just put here shirts costs $5 in pants, 10.

Generally what you want to be doing is you want to be connecting your data store to this or here in settings, you can have rag retrieval, augmented generation.

It's basically able to pull in whatever data you have, like the pricing on the shirts and the pants or whatever.

Then we're asking you to use a tool.

The tool is called Code Interpreter, right?

As you can see here, it's one of the extensions that's available within this vertex AI AI builder, AI agents builder.

Code Interpreter, similar to what OpenAI has, it can do stuff with calculations and numbers and stuff like that.

This code agent will use that to calculate the subtotal of sales tax and the total.

And then it's going to return to the other agent.

This isn't quite how you would actually do that.

I realized now right off the bat, I'm going to tell you that this isn't working too well right out of the box for me, at least.

And I think the reason for that is you really have to go here to examples and provide, well examples.

These are effectively few shot prompt examples for the LLM.

You describe in detail how you want that interaction to go.

And then once you have those in there, it actually can work okay.

But really it seems like the ability for this to function effectively, like really falls hard on how good your examples are, how many examples you have.

While I was able to get this to work, it didn't work too well.

However, here's a pre-built DMV steering agent that they've built that has tons of examples, right?

Here they all are.

And as you can see here in the example info, they tell you how that conversation might go, right?

The person, the user, the customer says, I want to book an appointment at a field office and that triggers the action book appointment, right?

Book appointment is one of the actions, and they're using the API to do that.

And there's, as you can see a number of other examples in the goal, we're telling the agent that their goal is to collect the customer's request and route the customer to The correct service among the variety of online services available from the department of motor vehicles, DMV.

We're going to start by greeting the customer and telling them you can help with renewing the driver licenses, et cetera.

Do not attempt to help the customer directly, right?

We're saying that's not your job.

Don't do it.

Always transfer them directly to another topic.

And there's a couple of topics you can choose from.

You can send them to this agent, agent renew driver's license for booking appointments at local field offices, transfer to agent book appointments and for topics not related To the above topics or sensitive topics like fees and pricing transfer to agent fallback, which we'll likely call a human person.

Well, we'll check this out in a second.

And they've attached a number of tools like the DMV renew tool, the DMV book appointment tool and Code Interpreter, right?

And here are the four agents that we have, right?

In this example where they wanted to book an appointment, they called on book appointment, this action.

Let's take a look at that agent, right?

And here we're telling them that we're going to use this tool, DMV book appointment tool to respond to the customer with a list of offices and addresses.

Ask the customers to select one of the available offices and set the office through the tool appointment tool.

And if the customer wants something else, do agent fallback.

Let's reset this thing.

We're going to use GMI 1.0 as our test model to test this agent.

And we're going to say, hi, it says, hi, I'm your DMV assistant.

How can I help you?

I'm going to say book appointment.

And it says, I can help you book an appointment.

Notice here, it called the agent, the book appointment agent, right?

The original, like the starred one, the DMV steering, I said book appointment.

It redirected me to this agent, the book appointment agent and says, provide me with an address.

I'll provide my address here and click go.

No, right.

Can't find it.

It looks like here on the code, it just says URL example.com.

This is not actually hooked up to anything.

I guess that should have been pretty obvious, but this is the thing that would kind of provide that functionality.

If it was properly set up, you would find the appointment.

It would help you select a time slot, et cetera, et cetera.

And there's a couple other ones such as retail, for example, that's one of the prebuilt ones that we can test out.

A lot of them have to be set to global if you're using the generative playbook.

That's some specific thing that you have to set.

And again, I'm noticing here, so they're always transferring directly to another topic.

As soon as they Figure out what you want, they're handing you off to either search for a product or place an order.

And they're connected to the Google store API.

Manage your shopping cart place order and Code Interpreter is shut off.

These are the examples of how conversations can go.

For example, I want to buy a phone and they send you to the search product agent.

They're calling that sort of top level agent steering.

We're going to say I want to buy a phone and they search the Google store API.

Again, nothing's returning because they're just this is not real, not real functional code that they're using, but in that situation, it would search for those phones and Return the answers to you at the end of the day.

What does this mean?

Is it good?

Is it bad?

Is it useless?

Is it revolutionary?

Well, one, I think they definitely got a couple of things, right?

The ability to write in instructions

And then just type in this to throw it over to an agent.

This might look confusing if you're dealing with this for the first time, but at the end of the day, all this is, is just a sequence of characters.

And then you write out the name of the agent, kind of your magical incantation to summon that agent.

And then the agent name is whatever you make it out to be.

If I copy and paste this, right, basically this is the thing that you use to spin up another agent to answer that thing for you, right?

And then you type in the name of that agent and it's ready to go.

Same thing with the tool.

Instead of tool name, you put in, well, the name of the tool, whatever it is, blah.

And this is likely what the future looks like.

This is how a lot of us are going to be creating these agentic swarms with multiple agents.

Each one of them taught to execute certain specific tasks for us.

And then one sort of master agent that calls upon them to do it.

Of course, if somebody doesn't really know how to code, if they're not a web developer doing stuff like this, the Jason or whatever that is needed to be used, that might be A little bit complicated.

But again, a lot of these chatbots like the cloud three, like GPT-4, they are able to provide this for you to some degree and they're getting better at it.

For people that are coders, web developers, this will be pretty easy.

It seems like for people that are not coders, it might take a few hours of kind of messing around with it to try to learn how to do it.

It's not super complicated because the most the things that you really have to get right are these, the instructions and the other agents and stuff like that.

It sounds like you really have to fill out this example section with a lot of examples that spell out to the agent exactly what it's supposed to be doing.

The better this section is, the better it will be at answering those questions.

With that said, these are just fancier versions of chatbots with a little bit more functionality, a little bit more understanding.

This isn't quite the ChatGPT moment for AI agents.

However, looking at this, how this is kind of laid out, it's important to understand that from everything that I've been seeing in various other research papers from NVIDIA And OpenAI and Microsoft, what you're looking at here will largely be replacing a lot of the code that we've been using in the past.

Instead of just pages of code, a lot more of the things that we're going to be doing with computers will look like this.

A lot of it will be just natural language, us writing very specific instructions in kind of normal language about what we want the AI to do.

We're going to add little clips of code here and there when they're going to need to do something specific.

Or I think the next wave of that will be that it itself will be able to know which APIs it needs to call on or which webhooks to use, whatever, right?

It will kind of translate the language into those specific code snippets.

But the point is more and more I'm seeing this used in various AI research papers instead of code of what we used to think of as code.

And I think Google here built like a proto AI agent builder, a very basic rementory kind of beta first wave AI builder, but there's still quite a ways to go.

Anyways, with that said, thank you for watching.

I hope that was helpful.

I know some of you might think this was a bit of a let down, but the important thing is this space is moving forward fast.

If Google and others will continue improving and iterating, we're going to see some pretty cool stuff very soon here.

This, this is a step in the right direction.

With that said, my name is Wes Roth and thank you for watching.