Skip to content
Roboto Studio / evidence

youtube / video

Your first AI agent

Jono AlfordOriginal source ↗

Original video

About this video

Looking to build your first AI agent workflow? We share the stack that powers our current lead enrichment, and the open source Lead Agent repo from Vercel. Go build it, or ask us to build it for you.

Full transcript

Edited transcript from the original Roboto page

Hey, how's it going? Hope you're having a great day. I want to talk about workflows and agentic workflows and how you can actually integrate them inside of your website and very specifically the ones that are very high value with very low effort. Because this should be basically like a gateway drug for you getting into, you know, the agentic stuff and everything else. You've probably rolled your eyes by now, but it's very, very important. And I think you should be using it today because I think there's going to be a huge amount of companies that are going to dwarf the efforts of other companies with being able to basically leverage AI.

So let's start. What do we currently use and what is valuable with our particular business? Well, we use a form enrichment system. So whenever somebody submits a form on our website, it takes the name, it takes the domain name that comes out of the email and it takes the message to understand intent and it creates a whole research chatbot that goes off and basically uses a bunch of tools to read the information from that particular client and come back and give me a complete synopsis of why they're getting in touch, whether they're a good fit, all these different informations, you know, financials, whether it's a series A round, series B round, all that kind of stuff. So that's really really valuable.

Now, how do you do that? And how would you get that running as fast as possible? Well, you're probably going to be using something like HubSpot or Salesforce. HubSpot is an interesting one. Depending on which plan you have, but there are a whole bunch of different trade-offs that are involved in there, whereas Salesforce is very very easy to do this with. But basically, you would stick it in between those two endpoints and you would make sure that it basically runs.

Now, how do you know this is going to run consistently? Because that is a problem of its own. Well, I'll tell you about the most underutilized thing that I know of that Vercel has, which is this tool called workflows or the use workflow development kit. It's right here. Now, what does make any TypeScript function durable mean? I don't know when I first looked at this, nor did anybody on my team know. Let me tell you what this actually means, because it's so much simpler when you know that it basically means that no matter what happens, this generally doesn't break. It will take another go at whatever it is doing.

So in the particular context that we're talking about here, if for example the AI responded with an incorrect system, so like it basically gave you an error when it came back, it would try again maybe three times before it basically failed. Or for example, it tried going to an API and the API wasn't available at that particular time. It can just basically go again and wait and go again. Even more impressive is that it can wait for a tremendous amount of time. So you could very easily say, "Go ping this after 5 minutes. Go ping this after 10 minutes. Okay, that didn't work. Go ping this after 3 weeks." It's really that functional. So like it will go that length of time.

Now, if you're a developer, you've probably just heard what I've said and thought, "Oh, wow. How do they keep it running for three weeks or a month or six months?" I don't know. Frankly, I don't know how they do it. It's magic. It works though and it's really, really incredible. And you can do some incredible onboarding flows with this particular technique. It's just that people aren't using it right now. And I think these companies that are going to use it are going to absolutely blow other companies out of the water.

So this is just basically some of the code. We can give some advice as to how to build these things for the first time. But the best way to just start working with it is to use something that somebody else has built, which is the lead agent. So this thing right here is something that Vercel has open-sourced and it is a perfect example very similar to what we were talking about where it is an inbound lead qualification tool that uses your own data set. So you can actually put your own data warehouse in here or whatever tools you're using to enrich that particular lead that comes through.

Now, this is a pretty big form. And I'm not the biggest fan of having to fill out quite substantial forms, but this shows you how the architecture works. And you can see, you know, user submits form. This starts the workflow. So, again, the workflow development kit here. And then the research agent goes off using AI's SDK. Again, the AI's SDK qualify lead. You can have your own thing. And this is the reason why it's so powerful. Once you've built this functionality, you can tweak it exactly to your business rather than having to use predefined tools that already exist that are kind of like a one-size-fits-all. Think of it like tailoring a shirt that you know is going to look really really nice versus buying one off of the rail and it kind of looks all right, but it's a bit boxy and it looks a bit weird. That's what we're currently used to with all these different SaaS solutions. This is the custom-tailored looking really really nice when you actually wear it.

So yeah, that's the difference here. And again, this is just using the Vercel architecture. I would really heavily advise using Vercel. Not because I'm getting paid for this, but because we just use it. And every time we use Vercel, there is no headaches with anything that we build. It's just really, really efficient.

And one thing that's really particularly nice about this demo in particular is the Slack bot that comes involved in it. So the Slack uses Bolt, which is a brand new library for Slack. I think it was released like three months ago from memory. But it's a way of Vercel basically putting this adapter that sits between those two toolings. I'm going to check this. I'm going to fact-check that this is a new tool from Slack, but I'm fairly certain it is.

And you might be wondering with all of this, this is all sounding great so far. But surely putting our conversions and things like that inside of here is a very easy way for data to leak and we should start worrying about that kind of like that information going out. Well, that is where Sandbox SDK comes in. So this in particular is usually used for if people are going to try executing code and doing things and trying to stop people from wiping stuff. But there is a whole host of different ways to ensure that your data doesn't leak from this AI agent. There are a lot of them built in. Vercel basically builds into this whole tooling. But this is one that if you want to do more complicated stuff is very very valuable.

And now I'm hoping your brain's spinning with a whole bunch of different things that you might want to do with your own form. But if not, I'll give you a couple more suggestions that are very very useful that I would consider doing right now and we are doing internally for different clients and for ourselves.

So internationalization is a massive one. Internationalization used to be so so expensive to do, and now it's gone down to pennies. So you can use tools like Languin, which is fantastic to be able to pass it back in markdown or pass it back in a whole different range of different file types, or you can integrate with AI itself to just rewrite pages and give you a great first draft.

And talking of first drafts, there's another option that you can do with this, which is if you're trying to do content velocity, you know, you need to put out blogs and you find yourself kind of bottlenecked by the number of marketers that you have on your team. Well, the great thing you can do is you can actually generate on a cron job. So when it goes off like once every week, it fires five different drafts off based on something like, say for example in our particular field, you might do it based on X posts like X kind of trending things or something like Hacker News. You can base those and create a bunch of different blog posts. So you can generate the first draft of say five blog posts every week at 9:00 a.m. UK or 9:00 a.m. US, whatever. That's something you can do with this particular tool. You can feed that into Sanity or Contentful or any other headless content management system that allows you to make API calls.

And then finally, probably another one that's really really useful that we see, some people are offering this as a SaaS, but you might want it internally, which is you might want to create an internal chatbot from the company information wherever that is kept. If that's in a data warehouse or if that's kept in, for example, markdown or it's kept inside of Sanity or an API. Getting all that information together and being able to query that information fast is like the fastest way to onboard new people to the team. That's all possible with everything that we've just talked about.

So if any of these things sound particularly interesting, get in touch. We're going to do a few more demos in the future where we talk about how to actually build these and show a lot more development environments. So if you like this, do us a favor. Give us the thumbs up or subscribe or whatever it is that people usually say at the end of these videos. Anyway, thank you ever so much for listening. Take care. Have a great day. Bye-bye.

Read timestamped source captions

0:04Hey, how's it going? Hope you're having

0:04Hey, how's it going? Hope you're having a great day. I want to talk about

0:07a great day. I want to talk about

0:07a great day. I want to talk about workflows and agentic workflows and how

0:09workflows and agentic workflows and how

0:09workflows and agentic workflows and how you can actually integrate them inside

0:10you can actually integrate them inside

0:10you can actually integrate them inside of your website and very specifically

0:13of your website and very specifically

0:13of your website and very specifically the ones that are very high value with

0:15the ones that are very high value with

0:15the ones that are very high value with very low effort. Um, because this should

0:17very low effort. Um, because this should

0:17very low effort. Um, because this should be basically like a gateway drug for you

0:19be basically like a gateway drug for you

0:19be basically like a gateway drug for you getting into, you know, the agentic

0:20getting into, you know, the agentic

0:20getting into, you know, the agentic stuff and everything else. You've

0:22stuff and everything else. You've

0:22stuff and everything else. You've probably rolled your eyes by now, but

0:23probably rolled your eyes by now, but

0:23probably rolled your eyes by now, but it's very, very important. And I think

0:25it's very, very important. And I think

0:25it's very, very important. And I think you should be using it today because I

0:27you should be using it today because I

0:27you should be using it today because I think there's going to be a huge amount

0:28think there's going to be a huge amount

0:28think there's going to be a huge amount of companies that are going to dwarf the

0:30of companies that are going to dwarf the

0:30of companies that are going to dwarf the efforts of other companies um with being

0:33efforts of other companies um with being

0:33efforts of other companies um with being able to basically leverage AI. Um so

0:36able to basically leverage AI. Um so

0:36able to basically leverage AI. Um so let's start. What do we currently use

0:38let's start. What do we currently use

0:38let's start. What do we currently use and what is valuable with our particular

0:40and what is valuable with our particular

0:40and what is valuable with our particular business? Well, we use a form enrichment

0:43business? Well, we use a form enrichment

0:43business? Well, we use a form enrichment um system. So whenever somebody submits

0:46um system. So whenever somebody submits

0:46um system. So whenever somebody submits a form on our website, it takes the

0:47a form on our website, it takes the

0:47a form on our website, it takes the name, it takes the domain name that

0:49name, it takes the domain name that

0:49name, it takes the domain name that comes out of the email and it takes the

0:51comes out of the email and it takes the

0:51comes out of the email and it takes the message to understand intent and it

0:54message to understand intent and it

0:54message to understand intent and it creates a whole research box bot that

0:56creates a whole research box bot that

0:56creates a whole research box bot that goes off and basically uses a bunch of

0:58goes off and basically uses a bunch of

0:58goes off and basically uses a bunch of tools to read the information from that

1:00tools to read the information from that

1:00tools to read the information from that particular client and come back and give

1:03particular client and come back and give

1:03particular client and come back and give me a complete synopsis of um why they're

1:06me a complete synopsis of um why they're

1:06me a complete synopsis of um why they're getting in touch, whether they're a good

1:07getting in touch, whether they're a good

1:07getting in touch, whether they're a good fit, all these different informations,

1:09fit, all these different informations,

1:09fit, all these different informations, you know, financials, whether it's a

1:11you know, financials, whether it's a

1:11you know, financials, whether it's a series A round, series B round, all that

1:13series A round, series B round, all that

1:13series A round, series B round, all that kind of stuff. Um, so that's really

1:15kind of stuff. Um, so that's really

1:15kind of stuff. Um, so that's really really valuable. Now, how do you do

1:17really valuable. Now, how do you do

1:18really valuable. Now, how do you do that? And how would you get that running

1:19that? And how would you get that running

1:19that? And how would you get that running as fast as possible? Um, well, you're

1:21as fast as possible? Um, well, you're

1:22as fast as possible? Um, well, you're probably going to be using something

1:23probably going to be using something

1:23probably going to be using something like HubSpot or Salesforce. Um, there

1:25like HubSpot or Salesforce. Um, there

1:25like HubSpot or Salesforce. Um, there are trade-offs with HubSpot. So, HubSpot

1:27are trade-offs with HubSpot. So, HubSpot

1:27are trade-offs with HubSpot. So, HubSpot is an interesting one. Um, depending on

1:29is an interesting one. Um, depending on

1:29is an interesting one. Um, depending on which plan you have, but there are a

1:31which plan you have, but there are a

1:31which plan you have, but there are a whole bunch of different trade-offs that

1:32whole bunch of different trade-offs that

1:32whole bunch of different trade-offs that that are involved in there, whereas

1:34that are involved in there, whereas

1:34that are involved in there, whereas Salesforce is very very easy to do this

1:36Salesforce is very very easy to do this

1:36Salesforce is very very easy to do this with. Um, but basically, you would stick

1:38with. Um, but basically, you would stick

1:38with. Um, but basically, you would stick it in between those two um endpoints and

1:41it in between those two um endpoints and

1:41it in between those two um endpoints and you would make sure that it basically

1:42you would make sure that it basically

1:42you would make sure that it basically runs. Now, how do you know this is going

1:45runs. Now, how do you know this is going

1:45runs. Now, how do you know this is going to run consistently? Because that is a

1:47to run consistently? Because that is a

1:47to run consistently? Because that is a problem of its own. Well, I'll tell you

1:49problem of its own. Well, I'll tell you

1:49problem of its own. Well, I'll tell you about the most underutilized thing that

1:51about the most underutilized thing that

1:51about the most underutilized thing that I know of that Versel has, which is this

1:53I know of that Versel has, which is this

1:54I know of that Versel has, which is this tool called workflows or the use

1:56tool called workflows or the use

1:56tool called workflows or the use workflow uh development kit. It's right

1:58workflow uh development kit. It's right

1:58workflow uh development kit. It's right here. Now, what does make any Typescript

2:02here. Now, what does make any Typescript

2:02here. Now, what does make any Typescript function durable mean? I don't know when

2:05function durable mean? I don't know when

2:05function durable mean? I don't know when I first looked at this, nor did anybody

2:07I first looked at this, nor did anybody

2:07I first looked at this, nor did anybody on my team know. Let me tell you what

2:09on my team know. Let me tell you what

2:09on my team know. Let me tell you what this actually means. uh because it's so

2:11this actually means. uh because it's so

2:12this actually means. uh because it's so much simpler when you know that it

2:13much simpler when you know that it

2:13much simpler when you know that it basically means that no matter what

2:15basically means that no matter what

2:15basically means that no matter what happens this generally doesn't break it

2:18happens this generally doesn't break it

2:18happens this generally doesn't break it will take another go at whatever it is

2:20will take another go at whatever it is

2:20will take another go at whatever it is doing. So in the particular context that

2:22doing. So in the particular context that

2:22doing. So in the particular context that we're talking about here, if for example

2:25we're talking about here, if for example

2:25we're talking about here, if for example um the AI responded with a uh incorrect

2:28um the AI responded with a uh incorrect

2:28um the AI responded with a uh incorrect system, so like it basically gave you an

2:30system, so like it basically gave you an

2:30system, so like it basically gave you an error when it came back, um it would try

2:33error when it came back, um it would try

2:33error when it came back, um it would try again maybe three times before it

2:34again maybe three times before it

2:34again maybe three times before it basically failed. Um or for example, it

2:38basically failed. Um or for example, it

2:38basically failed. Um or for example, it tried going to an API and the API wasn't

2:40tried going to an API and the API wasn't

2:40tried going to an API and the API wasn't available at that particular time. It

2:42available at that particular time. It

2:42available at that particular time. It can just basically go again and wait and

2:45can just basically go again and wait and

2:45can just basically go again and wait and go again. Um, even more impressive is

2:48go again. Um, even more impressive is

2:48go again. Um, even more impressive is that it can wait for a tremendous amount

2:50that it can wait for a tremendous amount

2:50that it can wait for a tremendous amount of time. So, you could very easily say,

2:53of time. So, you could very easily say,

2:53of time. So, you could very easily say, "Go ping this after 5 minutes. Go ping

2:55"Go ping this after 5 minutes. Go ping

2:55"Go ping this after 5 minutes. Go ping this after 10 minutes. Okay, that didn't

2:57this after 10 minutes. Okay, that didn't

2:58this after 10 minutes. Okay, that didn't work. Go ping this after 3 weeks." It's

3:00work. Go ping this after 3 weeks." It's

3:00work. Go ping this after 3 weeks." It's really that functional. So, like it will

3:03really that functional. So, like it will

3:03really that functional. So, like it will go that length of time. Now, if you're a

3:05go that length of time. Now, if you're a

3:05go that length of time. Now, if you're a developer, you've probably just heard

3:06developer, you've probably just heard

3:06developer, you've probably just heard what I've said and thought, "Oh, wow.

3:09what I've said and thought, "Oh, wow.

3:09what I've said and thought, "Oh, wow. How do they keep it running for three

3:11How do they keep it running for three

3:11How do they keep it running for three weeks or a month or six months?" Um, I

3:14weeks or a month or six months?" Um, I

3:14weeks or a month or six months?" Um, I don't know. Frankly, I don't know how

3:16don't know. Frankly, I don't know how

3:16don't know. Frankly, I don't know how they do it. It's magic. It works though

3:18they do it. It's magic. It works though

3:18they do it. It's magic. It works though and it's really, really incredible. And

3:19and it's really, really incredible. And

3:19and it's really, really incredible. And you can do some incredible onboarding

3:21you can do some incredible onboarding

3:21you can do some incredible onboarding flows with this particular technique.

3:23flows with this particular technique.

3:23flows with this particular technique. It's just that people aren't using it

3:24It's just that people aren't using it

3:24It's just that people aren't using it right now. Um, and I think these

3:26right now. Um, and I think these

3:26right now. Um, and I think these companies that are going to use it are

3:28companies that are going to use it are

3:28companies that are going to use it are going to absolutely blow other companies

3:30going to absolutely blow other companies

3:30going to absolutely blow other companies out of the water. Um, so this is just

3:33out of the water. Um, so this is just

3:33out of the water. Um, so this is just basically some of the code. Uh, we can

3:34basically some of the code. Uh, we can

3:34basically some of the code. Uh, we can give some advice as to how to build

3:36give some advice as to how to build

3:36give some advice as to how to build these things for the first time. Um, but

3:38these things for the first time. Um, but

3:38these things for the first time. Um, but the best way to just start working with

3:40the best way to just start working with

3:40the best way to just start working with it is to use something that somebody

3:42it is to use something that somebody

3:42it is to use something that somebody else has built, which is the lead agent.

3:45else has built, which is the lead agent.

3:45else has built, which is the lead agent. So, this thing right here is something

3:47So, this thing right here is something

3:47So, this thing right here is something that Versel has open sourced and it is a

3:49that Versel has open sourced and it is a

3:49that Versel has open sourced and it is a perfect example very similar to what we

3:51perfect example very similar to what we

3:51perfect example very similar to what we were talking about um where it is a

3:53were talking about um where it is a

3:54were talking about um where it is a inbound lead qualification tool that

3:56inbound lead qualification tool that

3:56inbound lead qualification tool that uses your own data set. So, you can

3:58uses your own data set. So, you can

3:58uses your own data set. So, you can actually put your own data warehouse in

4:00actually put your own data warehouse in

4:00actually put your own data warehouse in here or whatever tools you're using um

4:02here or whatever tools you're using um

4:02here or whatever tools you're using um to enrich that particular lead that

4:05to enrich that particular lead that

4:05to enrich that particular lead that comes through. Now, this is a pretty big

4:07comes through. Now, this is a pretty big

4:07comes through. Now, this is a pretty big form. Um, and I'm not the biggest fan of

4:09form. Um, and I'm not the biggest fan of

4:09form. Um, and I'm not the biggest fan of having to fill out quite substantial

4:11having to fill out quite substantial

4:11having to fill out quite substantial forms, but this shows you how the

4:12forms, but this shows you how the

4:12forms, but this shows you how the architecture works. And you can see, you

4:14architecture works. And you can see, you

4:14architecture works. And you can see, you know, user submits form. This starts the

4:16know, user submits form. This starts the

4:16know, user submits form. This starts the workflow. So, again, the workflow

4:18workflow. So, again, the workflow

4:18workflow. So, again, the workflow development kit here. Um, and then the

4:20development kit here. Um, and then the

4:20development kit here. Um, and then the research agent goes off using AIS SDK.

4:23research agent goes off using AIS SDK.

4:23research agent goes off using AIS SDK. Um, again, the AIS SDK qualify lead. You

4:26Um, again, the AIS SDK qualify lead. You

4:26Um, again, the AIS SDK qualify lead. You can have your own thing. And this is the

4:28can have your own thing. And this is the

4:28can have your own thing. And this is the reason why it's so powerful. Once you've

4:30reason why it's so powerful. Once you've

4:30reason why it's so powerful. Once you've built this functionality, you can tweak

4:32built this functionality, you can tweak

4:32built this functionality, you can tweak it exactly to your business rather than

4:34it exactly to your business rather than

4:34it exactly to your business rather than having to use predefined uh tools that

4:37having to use predefined uh tools that

4:37having to use predefined uh tools that already exist that are kind of like a

4:38already exist that are kind of like a

4:38already exist that are kind of like a one-sizefits all. It's the difference

4:40one-sizefits all. It's the difference

4:40one-sizefits all. It's the difference between tailoring a shirt that you know

4:42between tailoring a shirt that you know

4:42between tailoring a shirt that you know is going to look really really nice

4:44is going to look really really nice

4:44is going to look really really nice versus buying one off of the rail and it

4:46versus buying one off of the rail and it

4:46versus buying one off of the rail and it kind of looks all right, but it's a bit

4:47kind of looks all right, but it's a bit

4:47kind of looks all right, but it's a bit boxy and it looks a bit weird. That's

4:49boxy and it looks a bit weird. That's

4:49boxy and it looks a bit weird. That's that's what we're currently used to with

4:51that's what we're currently used to with

4:51that's what we're currently used to with all these different SAS solutions. This

4:53all these different SAS solutions. This

4:53all these different SAS solutions. This is the customtailored looking really

4:55is the customtailored looking really

4:55is the customtailored looking really really when you actually wear it. Um, so

4:58really when you actually wear it. Um, so

4:58really when you actually wear it. Um, so yeah, that's the that's the difference

4:59yeah, that's the that's the difference

4:59yeah, that's the that's the difference here. And again, this is just using the

5:01here. And again, this is just using the

5:01here. And again, this is just using the Versell architecture. I would really

5:03Versell architecture. I would really

5:03Versell architecture. I would really heavily advise using Versell. Um, not

5:06heavily advise using Versell. Um, not

5:06heavily advise using Versell. Um, not because I'm getting paid for this, but

5:08because I'm getting paid for this, but

5:08because I'm getting paid for this, but because we we just use it. And every

5:10because we we just use it. And every

5:10because we we just use it. And every time we use Versell, there is no

5:11time we use Versell, there is no

5:11time we use Versell, there is no headaches with anything that we build.

5:13headaches with anything that we build.

5:13headaches with anything that we build. Um, it's just really, really efficient.

5:15Um, it's just really, really efficient.

5:16Um, it's just really, really efficient. Um, and one thing that's really

5:17Um, and one thing that's really

5:17Um, and one thing that's really particularly nice about this demo in

5:18particularly nice about this demo in

5:18particularly nice about this demo in particular is the Slack bot that comes

5:21particular is the Slack bot that comes

5:21particular is the Slack bot that comes involved in it. So, the Slack uses uh

5:23involved in it. So, the Slack uses uh

5:23involved in it. So, the Slack uses uh Bolt, which is a brand new uh thing uh

5:26Bolt, which is a brand new uh thing uh

5:26Bolt, which is a brand new uh thing uh brand new brand new library um for

5:29brand new brand new library um for

5:29brand new brand new library um for Slack. I think it was released like

5:31Slack. I think it was released like

5:31Slack. I think it was released like three months ago from memory. Um but

5:33three months ago from memory. Um but

5:33three months ago from memory. Um but it's it's a way of versel basically

5:35it's it's a way of versel basically

5:35it's it's a way of versel basically putting this adapter that sits between

5:36putting this adapter that sits between

5:36putting this adapter that sits between those two toolings. Um I'm going to

5:38those two toolings. Um I'm going to

5:38those two toolings. Um I'm going to check this. I'm going to fact check that

5:39check this. I'm going to fact check that

5:39check this. I'm going to fact check that this is a new tool from Slack, but I'm

5:41this is a new tool from Slack, but I'm

5:41this is a new tool from Slack, but I'm fairly certain it is. Um and you might

5:43fairly certain it is. Um and you might

5:44fairly certain it is. Um and you might be wondering with all of this, this is

5:45be wondering with all of this, this is

5:45be wondering with all of this, this is all sounding great so far. Um but surely

5:48all sounding great so far. Um but surely

5:48all sounding great so far. Um but surely putting our conversions and things like

5:50putting our conversions and things like

5:50putting our conversions and things like that inside of here um is a very easy

5:52that inside of here um is a very easy

5:52that inside of here um is a very easy way for data to leak and we should start

5:54way for data to leak and we should start

5:54way for data to leak and we should start worrying about that kind of like that

5:56worrying about that kind of like that

5:56worrying about that kind of like that information going out. Um well that is

5:59information going out. Um well that is

5:59information going out. Um well that is where sandbox SDK comes in. Um so this

6:03where sandbox SDK comes in. Um so this

6:03where sandbox SDK comes in. Um so this in particular is usually used for if

6:05in particular is usually used for if

6:05in particular is usually used for if people are going to try executing code

6:06people are going to try executing code

6:06people are going to try executing code and doing things and trying to stop

6:08and doing things and trying to stop

6:08and doing things and trying to stop people from wiping stuff. Um but there

6:10people from wiping stuff. Um but there

6:10people from wiping stuff. Um but there is a whole host of different ways to

6:12is a whole host of different ways to

6:12is a whole host of different ways to ensure that your data doesn't leak from

6:14ensure that your data doesn't leak from

6:14ensure that your data doesn't leak from this AI agent. Um there are a lot of

6:17this AI agent. Um there are a lot of

6:17this AI agent. Um there are a lot of them built in um the Versell basically

6:19them built in um the Versell basically

6:19them built in um the Versell basically builds into this whole tooling. But this

6:21builds into this whole tooling. But this

6:21builds into this whole tooling. But this is one that if you want to do more

6:22is one that if you want to do more

6:22is one that if you want to do more complicated stuff is very very valuable.

6:25complicated stuff is very very valuable.

6:26complicated stuff is very very valuable. Um and now I'm hoping your brain

6:27Um and now I'm hoping your brain

6:27Um and now I'm hoping your brain spinning with a whole bunch of different

6:28spinning with a whole bunch of different

6:28spinning with a whole bunch of different things that you might want to do with

6:30things that you might want to do with

6:30things that you might want to do with your own form. Um but if not I'll give

6:33your own form. Um but if not I'll give

6:33your own form. Um but if not I'll give you a couple more suggestions that are

6:34you a couple more suggestions that are

6:34you a couple more suggestions that are very very useful that I would consider

6:36very very useful that I would consider

6:36very very useful that I would consider doing right now and we are doing

6:38doing right now and we are doing

6:38doing right now and we are doing internally um for different clients and

6:40internally um for different clients and

6:40internally um for different clients and for ourselves. So internationalization

6:43for ourselves. So internationalization

6:43for ourselves. So internationalization is a massive one. uh

6:45is a massive one. uh

6:45is a massive one. uh internationalization used to be so so

6:47internationalization used to be so so

6:47internationalization used to be so so expensive to do um and now it's gone

6:49expensive to do um and now it's gone

6:49expensive to do um and now it's gone down to pennies um so you can use tools

6:52down to pennies um so you can use tools

6:52down to pennies um so you can use tools like languin um which is fantastic to be

6:54like languin um which is fantastic to be

6:54like languin um which is fantastic to be able to pass it back in markdown or pass

6:56able to pass it back in markdown or pass

6:56able to pass it back in markdown or pass it back in a whole different range of uh

6:58it back in a whole different range of uh

6:58it back in a whole different range of uh different file types or you can

7:00different file types or you can

7:00different file types or you can integrate with AI itself to just rewrite

7:03integrate with AI itself to just rewrite

7:03integrate with AI itself to just rewrite uh pages and give you a great first

7:05uh pages and give you a great first

7:05uh pages and give you a great first draft. Um, and talking of first drafts,

7:08draft. Um, and talking of first drafts,

7:08draft. Um, and talking of first drafts, there's another option that you can do

7:09there's another option that you can do

7:09there's another option that you can do with this, which is if you're trying to

7:11with this, which is if you're trying to

7:11with this, which is if you're trying to do content velocity, you know, you need

7:12do content velocity, you know, you need

7:12do content velocity, you know, you need to put out blogs and you find yourself

7:14to put out blogs and you find yourself

7:14to put out blogs and you find yourself kind of uh bottlenecked by the number of

7:16kind of uh bottlenecked by the number of

7:16kind of uh bottlenecked by the number of marketers that you have on your team.

7:18marketers that you have on your team.

7:18marketers that you have on your team. Well, the great thing you can do is you

7:20Well, the great thing you can do is you

7:20Well, the great thing you can do is you can actually generate on a cron job. So

7:22can actually generate on a cron job. So

7:22can actually generate on a cron job. So when it goes off like once every week,

7:24when it goes off like once every week,

7:24when it goes off like once every week, it fires five different drafts off based

7:27it fires five different drafts off based

7:27it fires five different drafts off based on something like um say for example in

7:29on something like um say for example in

7:29on something like um say for example in our particular field um you might do it

7:32our particular field um you might do it

7:32our particular field um you might do it based on X posts like X um kind of

7:35based on X posts like X um kind of

7:35based on X posts like X um kind of trending things or something like Hacker

7:38trending things or something like Hacker

7:38trending things or something like Hacker News, you can base those and create a a

7:41News, you can base those and create a a

7:41News, you can base those and create a a bunch of different blog posts. So you

7:43bunch of different blog posts. So you

7:43bunch of different blog posts. So you can generate the first draft of say five

7:45can generate the first draft of say five

7:45can generate the first draft of say five blog posts every week at 9:00 a.m. UK or

7:48blog posts every week at 9:00 a.m. UK or

7:48blog posts every week at 9:00 a.m. UK or 9:00 a.m. US, whatever. Um, that's

7:509:00 a.m. US, whatever. Um, that's

7:509:00 a.m. US, whatever. Um, that's something you can do with this

7:51something you can do with this

7:52something you can do with this particular tool. You can feed that into

7:53particular tool. You can feed that into

7:53particular tool. You can feed that into Sanity or Contentful or any other

7:55Sanity or Contentful or any other

7:55Sanity or Contentful or any other headless content management system that

7:56headless content management system that

7:56headless content management system that allows you to make API calls. Um, and

7:59allows you to make API calls. Um, and

7:59allows you to make API calls. Um, and then finally, probably another one

8:01then finally, probably another one

8:01then finally, probably another one that's really really useful that we see

8:03that's really really useful that we see

8:03that's really really useful that we see um, some people are offering this as a

8:04um, some people are offering this as a

8:04um, some people are offering this as a SAS, but you might want it internally,

8:07SAS, but you might want it internally,

8:07SAS, but you might want it internally, which is you might want to create an

8:08which is you might want to create an

8:08which is you might want to create an internal chatbot from the company

8:10internal chatbot from the company

8:10internal chatbot from the company company information wherever that is

8:12company information wherever that is

8:12company information wherever that is kept. If that's in a data warehouse or

8:14kept. If that's in a data warehouse or

8:14kept. If that's in a data warehouse or if that's kept in, for example, markdown

8:16if that's kept in, for example, markdown

8:16if that's kept in, for example, markdown or it's kept inside of sanity or an API.

8:18or it's kept inside of sanity or an API.

8:18or it's kept inside of sanity or an API. Um, getting all that information

8:20Um, getting all that information

8:20Um, getting all that information together and being able to query that

8:22together and being able to query that

8:22together and being able to query that information fast is like the fastest way

8:25information fast is like the fastest way

8:25information fast is like the fastest way to onboard new people to the team.

8:27to onboard new people to the team.

8:27to onboard new people to the team. That's all possible with everything that

8:29That's all possible with everything that

8:29That's all possible with everything that we've just talked about. Um, so if any

8:31we've just talked about. Um, so if any

8:31we've just talked about. Um, so if any of these things sound particularly

8:32of these things sound particularly

8:32of these things sound particularly interested, get in touch. We're going to

8:34interested, get in touch. We're going to

8:34interested, get in touch. We're going to do a few more demos in the future where

8:36do a few more demos in the future where

8:36do a few more demos in the future where we talk about how to actually build

8:37we talk about how to actually build

8:38we talk about how to actually build these and show a lot more development

8:39these and show a lot more development

8:39these and show a lot more development environments. Um, so if you like this,

8:42environments. Um, so if you like this,

8:42environments. Um, so if you like this, do us a favor. Give us the uh the thumbs

8:44do us a favor. Give us the uh the thumbs

8:44do us a favor. Give us the uh the thumbs up or subscribe or whatever it is that

8:45up or subscribe or whatever it is that

8:45up or subscribe or whatever it is that people usually say at the end of these

8:47people usually say at the end of these

8:47people usually say at the end of these videos. Anyway, thank you ever so much

8:49videos. Anyway, thank you ever so much

8:49videos. Anyway, thank you ever so much for listening. Take care. Have a great

8:51for listening. Take care. Have a great

8:51for listening. Take care. Have a great day. Bye-bye.

Questions answered

Editorial answers from the original Roboto page.

What is a good first AI agent to build?

Lead enrichment is a great starting point. When someone submits a contact form, an agent can take the name, email domain, and message, run research across the web, and return a summary covering company size, funding stage, and likely intent. The value is high, the scope is contained, and it slots into a workflow you already have.

Read the original ↗
What is the Vercel Workflow SDK?

It's a toolkit that makes any TypeScript function durable. If an API call fails or an AI model returns an error, the workflow retries automatically rather than dying silently. It can also wait for long periods between steps, so you can build flows that check in after 5 minutes, 3 hours, or even weeks without keeping a server alive the whole time.

Read the original ↗
What is the Vercel Lead Agent?

An open-source inbound lead qualification tool from Vercel. A user submits a form, the workflow kicks off, and a research agent enriches the lead using your own data and tools. It includes a Slack integration via Bolt so results land in a channel rather than just an email. It's a working reference implementation you can fork and adapt to your own business logic.

Read the original ↗
What other agentic workflows are worth building?

Three worth doing now are automated content drafting (a cron job that generates blog post drafts from trending topics each week and pushes them into your CMS), internationalization (passing pages to a translation service or LLM to produce first drafts at a fraction of the old cost), and an internal knowledge chatbot that queries your company data warehouse or CMS so new team members can get answers fast.

Read the original ↗
How do you stop data leaking from an AI agent?

Vercel's Sandbox SDK is designed for this. It isolates execution so the agent can't reach data it shouldn't, which matters when agents are running code or accessing internal systems. Vercel also builds additional data guardrails into the workflow tooling itself.

Read the original ↗

Scope and context

  • Video authors are credited by the original Roboto page; individual transcript speakers remain unassigned.