Ilan (00:00)
random funny story, I was supposed to have a much brighter light in my office, but it hung too low and I kept hitting my head on it. So I ended up moving it elsewhere in the house ⁓ and then took a dimmer light from my office.

David (00:07)
⁓

There's some sort of Icarus joke in there.

Ilan (00:15)
I definitely

flew too close to the office light. my little wax wings.

David (00:19)
to the LED.

Ilan (00:24)
Welcome to Prompt and Circumstance. I'm Ilan

David (00:26)
And I'm David.

Ilan (00:27)
And today we're talking about AI agents.

on today's episode,

gonna tell you what is the definition of an AI agent, give some examples, and then we'll even build an And our goal here really today is to show you that

You could definitely be building them already without much new learning needed.

Ilan (00:57)
Now is a great time to talk about our first sponsor for this week,

Querio is a powerful tool for getting insights into your products. It allows you to ask natural language questions and then it will find the answers in your data. It saves hours a week for me, the product team I work with and our data team. And just for our listeners, they'll give you a two month free trial if you sign up today and let them know that David and Ilan sent you.

Ilan (01:27)
With that, let's get right into it. So when somebody says that there's a new agent, they have a new agent that they've built, what comes to mind for you?

David (01:36)
something autonomous, I'd imagine. Although, you I might be a little bit skeptical because that term gets, is getting overused more and more each day.

Ilan (01:47)
Mm-hmm.

I think that we should be using it even more, that there is an air of mystery that's been built around an agent, what an agent is, what it does. I know that you launched.

an feature for the company that you work What is that? Like, what went into that, if you can share?

David (02:08)
Sure, yeah, I'd love to. So calling it agentic AI as opposed to an AI agent at the moment. ⁓ And that might be, you know, a little bit nitpicky on my part. ⁓ But the way that I'm thinking about the distinction is that an AI agent really can act ⁓ autonomously in ways that might have been unplanned previously.

Ilan (02:15)
Mm-hmm.

David (02:35)
Whereas something that is agentic is something that ⁓ acts on behalf of the user, it goes through a fairly rigid or scripted ⁓ set of double click on what this is that we made, a sales

when they interact with their CRM, they don't like to do paperwork, right? They don't like to ⁓ log, say, sales activities. And this is something that I've heard from these outside sales reps who would say things like, you want me to do paperwork or do want me to sell? that comes into conflict, of course, with what the business wants, which is we want to have that record in case

you leave, let's say, we want to be able to continue that deal. Additionally, we want to have visibility into what's going on. is that funnel? Right? Let's get some confidence in that.

So,

this solves that problem for both. It makes it very easy for rep to log a sales activity So all that they need to do is use the mobile app for the CRM and talk at it and say,

I had lunch today with so-and-so from ABC Company. We talked about a new piece of machinery. And rather than having to tap, tap, tap and type into their phone, all of the context that is needed to create a sales activity is there. And so the application parses all of that out

pre-fills out a draft of a sales activity the rep to say, that line was good or no, change this or that and away it goes. If it's good, then it'll get submitted to the system. So in that sense, it's agentic because it goes and looks up the customer, it looks up the contact, it looks up what are valid values. It does all of that on behalf of the sales rep.

Ilan (04:38)
So this is interesting I think this may end up being really an interesting conversation because I think we may have conflicting opinions about what is an agent. What would that and make turn it from agentic feature to an agent?

David (04:56)
Yeah, I would like to see ⁓ maybe suggestions of what to do coming from the agent. So right now, it has to be the rep saying log this It would be good for the agent to say, yeah, I'll log the activity. And by the way, I noticed that you haven't

visited this other customer nearby. ⁓ Maybe you should do that, right? So that would be one aspect of it. The other would be, of course, being able to simply do more than just a single piece of functionality there.

Ilan (05:38)
Got it. I see where you're coming from. I tend to think an agent as

system that can take some input and then has some kind of chain of events that happens from there. That it acts autonomously based on previous event that happened, which also was some kind of agentic capability.

So it's like maybe chaining together agentic capabilities makes an Does that jive for you?

David (06:18)
I mean, ⁓ yes, to an extent. ⁓ In my mind, I get really hung up on the word agent because it needs to have agency. So ⁓ for it to decide on what to do ⁓ based off of the context provided, rather than ⁓ taking orders. If it's taking orders, then it might as well be algorithmic almost. ⁓

Ilan (06:30)
Mm-hmm.

Mm-hmm.

David (06:46)
And so I think that's where I get hung up on. I mean, I like the fact that you use the word, you know, autonomous in there. When I think about it, the use of an MCP server really comes to mind a don't know and by when I say we, I mean,

say the application or ⁓ the agent does not know ahead of time what the possible things that it can do are. ⁓ It needs to go and consult something like an MCP server to discover what it does have access to, what it can do, what data has access to, and then decide with agency what to do.

Ilan (07:27)
So I'm going to provide a couple of examples here. And I want you to tell me agentic or agent.

David (07:35)
Sure. ⁓

Ilan (07:37)
summarizing an email inbox and then providing action items for a user. Is that an agent or is that just an agentic capability?

David (07:48)
put that closure to being an agent ⁓ because it provides the action items. ⁓ So, you know, if it were just summarizing an email inbox, that's agentic. And even then that might be a little bit of a stretch. Yeah.

Ilan (08:06)
All right, how about prepping docs before a call? Like for example, grabbing previous notes from the same chain or from the same set of invitees and maybe grabbing LinkedIn bios or maybe some a notion documents that you might have and then preparing like a summary doc.

David (08:27)
say that's more like an agent because ⁓ there are a number of different areas it needs to look and ⁓ it doesn't know what it has ahead of time. ⁓ And so it needs to make the decision and choose which one of those things to use and then ⁓ decide what to do with that information.

Right? mean, it's one thing to grab, hey, here's a scrape of the LinkedIn bios of the people who you're going to talk to. It's another to a reasonable summary or report ahead of time, a pre read that is relevant to the meeting that you're about to have.

Ilan (09:12)
All right. And how about getting relevant news an by user and then the user regular updates on what's going on.

David (09:25)
The fact that it decides on maybe when to provide the user with updates is on the agent side of things because again, it's acting with agency. It's making a choice as to whether or not to notify the user. So I put it into that camp.

Ilan (09:39)
Mm-hmm.

All right. I think we're a little closer on this than I feared when we started the conversation. was worried I was going to launch into my next section and you're going to be like,

David (09:49)
Hahaha

That's just agentic AI Yeah. Yeah. Just poo poo all over this. ⁓ No, think that was what? A four for four?

Ilan (10:06)
Three for three, yeah, three for three, well, then I think that we've agreed on a framework for what makes an agent. And as you said, the key piece is the ability to take in some information, not just from the user's input and then act on that information.

David (10:07)
Two for three, yeah, all right on.

in terms of our brief description of it. Yeah, definition of it. Yeah.

Ilan (10:34)
you hear a lot of terms getting thrown around. You you mentioned MCP, we have an episode about MCP servers, but you hear all kinds of terms like A2A or that get tied in.

with this concept of agents or building AI agents. And because of the jargon around I think that it makes people feel like there's a bigger barrier to entry to building agents. So I want to give a quick tutorial here on building an agent.

And I'm going to use that last example that we took, where I heard a little bit more hesitation from you, David. But I'm hoping by the end of this, I will have convinced you that we've created an agent together.

David (11:18)
Hahaha

Yeah. Hey, you know, don't get me wrong. I am no expert in AI agents. I simply with my words. So I would love to learn more about creating an AI agent.

Ilan (11:28)
haha

Ha

Cool. All right. So the example we're going to think through here is creating an that's going to scrape through relevant industry news for a is a problem that both you and I talked about in ⁓ some earlier episode where we are...

both in a situation where just keeping up with all of the news in our industries on top of the news for the podcast and on top of news for other areas of interest that we may have can be a bit daunting on top of other work that we need to have and spending a little time parenting and whatever else we might want to do with our time, right?

David (12:21)
Yeah, yeah. There's a lot stacked onto the plate. know, somebody other day asked me, hey, what sports do you follow? like, none. Sorry, I follow AI news. Do you want to talk about MPCP servers?

Ilan (12:37)
Did you hear about the windsurf acquisition?

David (12:39)
Yeah, right? Yeah.

André Karpathy, my man.

Ilan (12:46)
So this is, this is, ⁓ this is the problem that we're going to break down here and like, ⁓ a good product we should really start what's the, what's the flow, right? What happens for the user. And I think about this as a way to break down the steps that create the agent.

Ilan (13:07)
a great time for me to tell you other sponsor for this

Colab is a product that will help you and a you could run as a side is great for product managers or any professionals who are looking for a way set up their side business but don't really have enough time to do the initial market testing and

hand you a working product and they'll even find your first paying customers.

click on the link below and let them know that David and Ilan sent you for $250 off Validate and Build.

Ilan (13:39)
So let's think about what have to happen if you wanted to achieve this a And that actually breaks are the various agentic capabilities that need to be connected together to create the agent.

Again, we've talked about in the past, what is the value that product managers bring in the world of vibe coding and AI product creation? It's this, right? It's breaking down. What is the problem and how do you actually chain together actions for a user to create a flow ⁓ that comes from the understanding of what the user needs? So here's what I came up with.

the user has to information about what they're looking for, their role? What industry are they trying to understand And agent needs to define a so the first step is taking that and

David (14:34)
Mm-hmm.

Ilan (14:38)
kind of like what you described in the agent that you've created in your job, Fill out the relevant fields what you want that user to be able to or what you want the subsequent agents to be able to do from there. So what's the role? What's the are maybe the relevant industry publications that they would wanna...

be looking at where else would they want information? Maybe it's from X or LinkedIn. And how would you get that information?

David (15:07)
Mm-hmm.

Ilan (15:09)
actually search for those get some you think is the right information a relevant time you would make a decision about for this user. maybe you'd say, let me look at news from the last Maybe let me look at news from the last month.

David (15:25)
So this is where you would tell the agent or previously tell it what sources it has access to. Like you only have access to these news outlets, for example.

Ilan (15:40)
I don't think so. I don't think so. I think that this is a case where there are models out there, like that actually just go out and search the based on some API input.

David (15:54)
Okay.

Okay, so basically the tool that it would have access to is ⁓ an internet search. Yeah.

Ilan (16:02)
That's right. So

it needs to define what is a good internet search for ⁓ this user, this user's context.

So goes out, searches for relevant industry news, gets those sources, reads them, then builds a summary citations, a good bibliography, right? Something that the user will trust that it's not hallucinating information about the industry. Imagine if, you know, we set this to work on the vibe coding industry and it's like,

this new upstart podcast and Circumstance has shaken the whole industry, ⁓ likely unicorn valuation. While I would love to believe that, I would also know that that's probably that's not true, you know.

David (16:50)
Not yet, not yet anyway.

Ilan (16:52)
Not yet true, not yet true.

So create that summary believable by the user with citations, then deliver a preview, like a first version to the user that they can see, see what that looks like. The user gives a thumbs up, then it would set up some kind of a schedule to keep looking for news to deliver to this user in some way. Let's say to start.

through the UI, have to come back to the site maybe by email or through Slack or some other, you know, a push notification in the future. ⁓

David (17:30)
And

there would be another set of tools or resources that the agent would have access to. So if it has access to a Gmail integration or Slack integration, it needs to know about that.

Ilan (17:48)
Exactly. And this is where, you know, there's more of a UI UX component to it, where, know, you need to like allow the user to, to tell the agent what it has access to so that it can make those decisions.

David (18:03)
Mm-hmm.

Ilan (18:04)
So this is the flow. Then it has to keep track of time, right? So that it knows, hey, two days have passed. Two days is a big news cycle for the industry that this person is in. So probably a good time for me ⁓ to go look for some updates from the last couple of days and send the user an update versus, you know, my day job is in the logistics industry. Logistics moves a little slower. Probably a weekly update is like,

find a dandy for me.

David (18:34)
I see. Okay. So is it is it that the agent decides when to get an update? Or is it the user would set that

Ilan (18:45)
I think that...

In the scope of this agent, I still think that this is an agent, even if the user decides I want a weekly update versus a daily update. ⁓ but I think it's a more powerful agent. If you have the agent decide, Hey, I'm just constantly keeping tabs on what's going on out there in your industry. And I think now is a relevant time because some big news story.

David (18:59)
Okay.

Ilan (19:17)
happened, right? That might be the looking at the vibe coding industry, the Thursday to Monday of the windsurf end of the open AI deal to the hiring of the founders by Google to the acquisition of the remaining company by cognition.

you might want like every eight hour update if you're keeping track of that, where most of the time maybe every three or four days would be sufficient. So I think that would be like an upgraded or a future feature for this agent.

David (19:48)
Yeah.

Yeah, no, that's all right. mean, you know, this is solving a real problem. ⁓ And as you said, it's something that both of us feel as a problem, something that we both experience. So I'm looking forward to seeing this come to life.

Ilan (20:13)
Mm-hmm.

Alright.

So ⁓ having looked at our little Miro flow there.

thing to do is to really think through

What are the AI capabilities that are needed to chain together these steps in the flow? And I think here is where people get hung up, maybe underestimate what's possible from LLMs. So here's where I suggest taking a ⁓ user flow or

maybe an agent flow would be the right way ⁓ to call it. And putting it into an LLM. I have mentioned before, I like Claude these days, but you can use the LLM of your choice. Probably use the paid tier if you have it. And start to interact with it as an AI engineer, ⁓ prompt engineer.

and start to ask it about what do you think about my flow?

Am I missing anything? And how might we achieve this flow? Are there open AI or Anthropic or other models out there that we would use that you would suggest to use to, ⁓ to chain these actions together.

David (21:59)
Yeah, absolutely. I mean, using ⁓ LLMs as a thought partner, that's something that was brought up in the panel discussion as well. So certainly a very compelling use case.

Ilan (22:11)
And one thing I also suggest here, just from personal experience is try throwing some of those random terms that I mentioned before, try throwing those at the LLM and just ask it is MCP or RAG or A2A a relevant, relevant step that I need to take here. And in this case, probably A2A, which stands for agent to agent.

David (22:35)
Mm-hmm.

Ilan (22:41)
would be something that's relevant. there is a standardized, there's becoming a standardized structure of how to pass information from one agent to another.

but you actually don't need to have that to make this agent come

David (22:56)
Yeah, this seems like it'd

be pretty good standalone.

Ilan (22:59)
Yeah, exactly. And then the last step is get it to deliver you a PRD for a junior engineer.

agent doesn't have to have a UI to serve its purpose, but building a microsite is a good way to test out your agent's capabilities, deliver it very quickly to users. So you could have a site, an app, an API, an MCP server, any one of these things, but with the rise of vibe coding and the commoditization of vibe coding.

that sprung up this podcast into being, ⁓ you know, I think that it's fitting to use one of those tools to actually try and build this up and deliver a functioning agent. So ⁓ from here, we can give that a try. How does that sound?

David (23:54)
That sounds amazing. That's a great connection to the overall theme.

Ilan (23:59)
All right, we ended up in Replit. Replit is great for code heavy activities. And I think that for building agents, it's the right type of tool where it might focus a little less on the UI of the feature, but focus more on that actual backend capability that's really delivering value for users.

So with that, we're gonna attach our markdown. We're going to tell Replit what to do,

David (24:30)
And ⁓ as Ilan is typing this out, if Replit does anything like it was ⁓ during our track meet, we're going to be cutting a whole bunch of time in between because Replit takes its time to work through everything.

Ilan (24:41)
Ha ha ha ha ha ha ha

David (24:45)
⁓ boom. Look at this.

Ilan (24:48)
Alright.

David (24:49)
Oh, I like that. That little 3D. Okay. So for those listening, we have a pretty nice looking page with some kind of 3D panel and he's trying to rotate it all the way around. Okay. All right. That's fun. That is, uh, that's entertaining. All right. Uh, I like this. It's,

I like the choice of ⁓ serif fonts by the way. That's very news oriented.

Ilan (25:23)
It

I did specifically tell it to use our ⁓

podcast as an inspiration ⁓ when I got Claude to generate the PRD. So I didn't get the colors right, but the serif fonts, yes.

Ilan (25:40)
All right, so let's take a look at what Replit created. So let's begin our journey start out by giving an email address.

Let's use one of sample descriptions over here and see what we got.

All right, look at that. So we've got the role analysis done. It shows some news categories for us. As we said, we wanna see industry news, competitive intelligence, functional news. It didn't select adjacent markets, because I guess it decided that for this healthcare related role, that wasn't so important, but it did choose regulatory and compliance. We could add additional keywords, some content length, frequency.

and it talked about some recommended sources. Then we're going to generate a preview.

All right, so we have this preview. It's in a markdown format that's not parsing the markdown format correctly, but we can see AI and healthcare technology regulatory landscape with an executive summary.

and we have some summaries within here it also talks about the source articles one from ⁓ may of this year one from june of this year and a couple that are a little bit

check, we'll make sure these are real articles and it gives you some insights about the discovery. So with that, you can see that the agent did work. It did create a new summary for us from what we assume are real articles that it found online.

Ilan (27:19)
all right. So as you saw there with some hiccups, maybe some, ⁓ aspects of the flow, we got an agent to at it. It connected.

the dots the way that we expected

So the point here really is if you are, about using agents in your solutions for users.

the path to get there is a lot cleaner and a lot easier than you may think when you just hear the term AI

David (27:51)
right. So at the end of the day, whether it's agentic AI or an AI agent, that only matters to word nerds like, like me. ⁓ I would say that as long as what this is solves a meaningful problem, in a way that works well for the business, you know, that that's really what this is about.

I think this is a great opportunity for people who are making changes to products or enhancements to products to think maybe from first principles as to, you know, given the tools that we have today, what if we weren't burdened by the existing UI or existing infrastructure that we have? Could we do something that would really solve this problem in a

completely different way and a completely like 10 X way for our customers.

Ilan (28:47)
So with that, I hope that this episode has demystified building AI agents or agentic AI for our listeners. I hope that you learned something. Leave us a comment. If you did, let us know what

you took away from this, if you've used an AI agent in your product, we want to hear those stories. Let us know what those are. I'm super interested to hear how our audience has been ⁓ leveraging these capabilities or maybe is planning to leverage these capabilities in the future. Otherwise,

David (29:27)
Yeah.

Yeah. Thanks for walking us through all of this, Ilan. It was a great learning experience.

Ilan (29:37)
Thank you. And with that, you can always find us on the socials at PNC podcast, leave us a rating or review, five stars, I hope, and we'll see you next time.

David (29:52)
See you next time.

All right.