Hi guys, I'm Alex.

So I'm going to be talking about how to build agents with no code.

In terms of background, I am not a developer.

I aspire to be one.

I have a lot of respect for people that have computer science degrees.

My background is more in finance.

But I spent the past six months playing around, tweaking with a lot of this different tech

that's coming down the pipe, and really enjoy it.

So every once in a while I come across something that's really interesting that I feel like

is worth sharing.

The thing that I think is worth sharing today is talking about some of the developments

that came off the back of the open AI dev day that happened a few weeks ago.

Now I feel like there wasn't enough focus and talk about this because at least a lot

of the entrepreneurs that I know in this space kind of looked at it as like, well, a

lot of the features and functions that they brought to market were kind of already there.

So if you actually look at what they brought, it was, OK, we're making GBT faster.

We're making it cheaper.

We're having a bigger context window.

But the big thing that was coming out was this introduction of agents.

And that's really what I'm going to focus on is, OK, what is an agent?

How do we build one?

And is it useful?

Now agents are, I'll get into, ultimately, a combination of a few things.

So it's a store of knowledge, retrieval.

It's the ability to then process that and take multiple steps.

And it's the ability to then call functions that sit outside of chat GBT.

And this is kind of a new thing that I feel like is now available to a lot of us that

are not developers now and actually makes it really useful.

So with that being said, I'm going to touch upon what are agents.

In this talk, we're going to try and build one from scratch using the chat GBT plus Zapier

AI.

I'll talk a little bit about what is Zapier and their recent launch of their AI product

and how these things kind of plug together to be really useful.

And ultimately, what we're going to build is basically a chat bot that we can be, more

or less, is kind of our centralized command center where we can then talk to it and ask

it to do stuff for us, and it'll do stuff for us in the real world.

So with that being said, I kind of talked a little bit about agents.

The key core thing here is that they're going to have a base of knowledge that they can

retrieve information from.

They're going to have a set of instructions that we're going to preset in terms of their

personality or in terms of things that they understand.

And then again, it comes back to the thing that I think is most interesting, which is

the ability to call functions.

So in the new UX or the UI setup, essentially on the left hand side, if you use chat GBT

or if you're a plus user, you should be presented with this piece up here, which essentially

allows you, if you click this button, explore, you can then start to configure and create

an agent.

So with that in mind, I'm going to jump straight in and see where we get to.

So thank you.

If we go to explore, there's a whole bunch of agents here that I've built previously.

If we go here and create an agent, so basically what we're going to be doing here is essentially

this brings up this tab where we can basically build a bot that will do stuff for us.

Now in here, we can configure the agent so we can set up what's the name, what we'd like

it to do and configure the three things that I talked about.

So in this case, let's just call this, I don't know, personal assistant GBT.

And so instructions, I'm going to leave that blank for now, but essentially we can set

out a number of instructions.

We can attach files to this.

So basically you can imagine if you wanted to build, let's say, a customer service agent

that has knowledge of your product, you could then attach here, you know, basically your

guidelines about your product, about your company here.

And then when you talk to this agent in the future, it's going to have knowledge and information

that it wouldn't have just by the data that's been trained on the web.

So here, one of the things that I do, and actually one of the things that really frustrates

me about chat GBT, and I appreciate I've misspelled this, but one of the things that really frustrates

me about chat GBT is how robotic it sounds when you ask it to write stuff for you.

And so as I tried to kind of address that one of the sort of hacks or get arounds that

I figured out was basically if you give it a whole bunch of content that you've written

in the past.

So you know, a bunch of emails just chunk them all into one prompt, and then essentially

give it a prompt that says, Hey, can you give me a detailed set of guidelines that a language

model could follow to sound like me.

And so I've done that for a number of different things.

So I've done that for my emails.

I've then, you know, figured, Well, we don't communicate the same way that we do.

Let's say writing board minutes as we do with like our mom.

So then I just started taking conversations with my mom during the same thing and conversations

of board minutes and doing the same thing.

So essentially I built a series of guidelines that set out different tones that, you know,

I can use in different contexts.

And so what I've done here is I've actually built those sets of tones into let's just

see it should be on the desktop.

A text file that is yeah, communication guidelines.

So communication guidelines will have different sets of things for different tones that I

want to use.

I'm not going to get into all the different bits of what's the tone of how I would like

to write and speak.

But what I will do is I will attach this now to the agent.

So basically what this is going to do is now the agent's going to have access to the guidelines

of, you know, essentially how I speak.

So if I go, let's say desktop communication guidelines, great.

Open for full-neck at spell.

Small assistant.

Okay.

So that's uploaded.

So this is basically step one is we've uploaded some knowledge to it.

I haven't put instructions.

I'm holding off on that for a reason.

And the other bit is I'm holding off on creating actions.

So before I do this, I can hit save.

Let's just go only me.

Great.

So if I ask this, hey, can you read?

Let's say, hey, what tones do you have access to?

Great.

So there should be three here.

One is basically what I'm calling my LinkedIn sort of, you know, tone, which is a set of

guidelines that how I would speak more or less business casual.

Another one is family and friends.

And then another one is going to be, you know, formal.

So essentially what I can do now is anytime I write an email, I can then say, hey, instead

of just saying that improve this and send it to chat, GPT, what I can say is, hey, improve

this with my LinkedIn tone.

And it'll automatically defer to that and link to that.

So that's step one.

Set two comes to instructions and comes to setting out the actions.

And those two things, I think, bring us down to the next step, which I think is fascinating.

Basically this business Zapier, which if you don't use Zapier, you should, it is a no-code

coding solution that basically lets you plug one thing to another thing.

So a very simple example could be, hey, whenever I finish a workout, put that workout into

a spreadsheet, or you can basically build very complex ones.

There's another business called Make.com, which is very similar.

So for me, it's the easiest way to kind of get up and running with an automation.

The challenge with doing this, though, is it is still a bit tedious.

It's not as tedious as going out and coding something up, but it is still a bit tedious

of sitting there and then saying, OK, I want X to do something, Y, and then getting all

the triggers and everything to line up.

And what I think is really fascinating is just around the time of this dev day when chat

GPT or OpenAI was moving to this model of agents, well, Zapier came out with this essentially

AI product, which was designed to link, you can basically link whatever agents you create

to any set of tools that are on the Zapier platform and automatically call those tools

in the your actions.

And we're going to do that now as we add that to the bot.

So that takes you to this page.

Essentially, there's two things that you need to copy, no coding required.

So if I come back to my personal assistant that we were just building, if I go here,

add a GPT, configure, I can drop down to actions, and now hit import from URL.

Simply by copying that one line in and hitting import, it's going to drop a whole bunch of

JSON stuff in here, which you don't really need to know what it's doing.

But essentially, it's now creating the actions that this bot can now reference to Zapier.

And Zapier has roughly 20,000 apps that are on Zapier, so you can basically pull in information

from lots of different places with that.

The second bit is the instructions.

So essentially, all you need to do is copy in this bit of instructions into the instructions

tab.

They've gone ahead and built a more detailed calendar bot that you can use.

So I'm going to just copy the full instructions from this calendar bot.

And that will be all we need to do in terms of creating this thing.

And then we can start playing with it and see what it's capable of.

So if I come back here, go GPT, edit, configure, instructions, update.

Great.

So now what this personal assistant GPT has access to is any other app that I decide to

flip on.

And the way I can flip those on, so for instance, so let's personal chat GPT should have access

to my calendar.

So now what I can do is say, hi, assistant, please help me prepare for the Mind Stone

event on my personal calendar tonight.

Now what it's going to do is basically going to take my requests in the request.

I'm basically asking it for something that is involved in my calendar tonight.

So it's automatically going to take that information and say, OK, do I have access to your calendar?

Should I go get access to your calendar?

And so it's now starting this action, which is running in the background.

And so here it's then going to prompt me, do I want to allow it access or not?

I'm going to hit allow.

And now it's going to Zapier and basically fetching that information.

And essentially with Zapier, it's going to pull up another screen that then allows me

to kind of toggle on what permissions I'm giving it access to.

So let me just jump into one where I ran it the first time and it worked.

So essentially it says help me research my next meeting.

Here it pulled up the correct meeting.

And then it basically, I said this is the event, so it came up with the event.

And then it asked me, you know, would I like to search more information about Mind Stone

or specific attendees related to this meeting?

All I said was yes, please.

And then it goes and searches the web and actually it found a bunch of information about Josh,

which was quite interesting that it talked about how Mind Stone's a learning platform.

It talked about their recent partnerships, product features, seed funding, background

and vision responses to the pandemic, a whole bunch of information that frankly I didn't

even know.

And what was fascinating then was then actually what I started doing was, and just for the

sake of time, I'll stick with this particular one.

But I then said, okay, great, let's just assume I've just gotten this bio of this person

that I'm now supposed to meet at this event.

You know, it's prepping me for an event that I'm going to.

I then just took my LinkedIn bio and downloaded that to PDF, dropped the PDF in here and said,

hey, here's my bio.

Can you please find some stuff that we might have some crossover about and could talk about?

And immediately it came back and says, okay, you're both interested in AI and technology,

you're both entrepreneurial experience, finance background, et cetera, et cetera.

And so it gets into these things and essentially what you can then do is the next step here

is because this bot now has access to whatever I want to give it to.

I could then basically get it to draft an email to Josh to reach out.

So for instance, I could say, okay, can you please put together a, let's say, compelling

and thoughtful email to Josh to reconnect after the conference.

He's used my, let's call it LinkedIn term.

And ultimately what it's going to do is it's going to basically build a draft and then

all I got to do is say, yes, please send that draft and it will then reach into my email

and send that out.

And so there's basically, I mean, there's a whole bunch of stuff you can do with this.

It's pretty endless in terms of the possibilities.

So I'm just going to let it finish.

And so basically what it does, and I've kind of skipped this step because I did it before

for the sake of time, is whenever the bot will then run into situations where it doesn't

have access to an app.

Let's say I'm asking it for some other piece of information that it doesn't yet have access

to, it will prompt me to come to this link and turn it on.

And so it'll say, you know, do I want to give it access to say my workouts or my calendar

or something else that's in my life.

And all I need to then do is then say, okay, sure, let's add a new action.

And we can basically add that on.

And so because I've already turned all these on, there's a whole bunch of other things

we can now do with this personal assistant bot if I come back to it.

So let's just say I am, so this is just a separate example, but let's say I'm planning

a dinner for let's say six AI entrepreneurs in London.

Please find, probably should just, on Monday from 6 p.m. to 8 p.m., please find me relevant

places which have featured on this blog.

Great, so I'm just going to take a blog here that's popular in London for interesting places

to go out.

And so now essentially what this is going to do is going to go and fetch the internet

for, you know, search the internet for this, pull out interesting places, give me some

recommendations.

And what I can do then is essentially get it to then create the calendar invite and invite

the people that I'd like to invite from.

So it's just a similar example, you know, both examples using calendar, but there are

many other examples that we can run through.

Sorry, let me just go here.

Okay, give me five.

Oops, it's always when doing a live demo, it works great, you know, the time before.

So I just find this is so fascinating because I've just discovered this yesterday and I've

been playing with it through half the day since.

And the example that I did just before I came over here, I did get it to create a calendar

invite and add the people that I suggested when I just copied in from a spreadsheet.

So that's basically the combination of agents in Zapier.

Those were two examples.

The other example, which I think I'm a little bit short on time, which so I won't get into,

but the other example I was working on earlier today, I was basically, you know, fetch the

last 10 Slack messages I've gotten, you know, evaluate if I need to respond to any of them.

If I do then, you know, draft a reply and create an Asana task for me to reply.

And then also, you know, link into my calendar and find time on my calendar to then schedule

time for me to then complete that Asana task and take care of it.

So I think as you start to then plug in many, many of these different things in different

places, I think this is kind of interesting to conclude on, you know, how this is moved

from just us engaging with chat GPT, kind of getting some information back or maybe

getting it to write a poem in a really interesting or weird way to then the second step was really

this introduction of agents, which if you've been around this space for the past six months,

there's been a whole bunch of them like agent runner and auto GPT and agent GPT that basically

were able to connect to the web, take multiple actions, plan those actions out and take steps

for us to now I think what's been the really big step in the past two weeks is been broad

competency, which which now we have agents that can basically access lots and lots of

different things in our life if you give them permission to, which there's, you know, questions

and dangers around that, sure.

But you can definitely see this sort of command center working for us to do lots of different

things. And instead of me jumping around from many different apps to just sit there in my

chat GPT terminal and then get it ask it to do different things for us. And I think where

this is going is basically these agents, if you start working with them and building with

them are going to increasingly, you know, move from things that we need to tell it every

time what we would like it to do to ultimately things that understand what we want to have

full autonomy to basically just go take actions on our behalf, you know, independently, which

I think is going to be really exciting and really fascinating time. And then I think where

this ultimately kind of ends up is where you have teams of agents or people have called

them swarms of agents, each with their own set of heuristics, their own set of access

to different bits of information, and the ability for maybe an agent to, you know, be

a supervisor of another agent. So you can imagine you have teams of agents, each with

their own specialty purpose, that can then go take, you know, much bigger actually and

strategic actions in the world. And I think that evolution from where we are right now,

right here, to having greater autonomy agents, I can see that playing out in the next six

months and I'd highly encourage you guys to get involved, play with the stuff, it's super

interesting. If you'd like to connect and just chat with me, this is my details. And

that's the end of my talk.