# Show Me Your AI — Episode 14

Full transcript: What are AI agents and how can you actually put them to work.

---

What even is an agent? And where does it live?

You know, he's like agentic workflows. Is that, like, secretly lots of people doing work in the background,

I think Ben is a secret agent.

Welcome to Show Me Your AI with Dan Gilbert.

We're here to talk about agentic workflows, but starting us off Ben Vincent as CTO to talk to us about, well, how are these agents and where do they actually live. Which I know is a question that the normies don't want to ask Ben. But I had it. And lots of people are wondering.

So talk us through this first, please. Absolutely. So where agents live isn't that different from where they where software lives? So when we worked on the left in a SaaS world, we had products that lived in the cloud, basically sitting on computers, running on lots of computers in big service centers. The way you'd access them is through some sort of interface.

So usually some sort of web interface. But essentially you're still talking to a piece of software in a data center. Now, if you look on the right in the world, the agents do the same thing. You are still running a bit of software in a data center, and the agent sits in a data center as well. Is just the way that you communicate with that.

Software is changed. So now you can work through Slack, through an email, through a web app or SaaS app, and you can talk via an agent. And then your agent does the part of talking to that software for you on your behalf. But essentially, it still all lives on the cloud, in a data center, on actual computers, on actual computers, ultimately on physical machines that literally live in data centers.

Like joking aside, like, it took me a while to get my head around this because people showing me agents, it was like, well, where is it? And I think the first explanation I had even half goes, well, it's anywhere you want it to be. Like, what is it like a unicorn? Like a magical, ethereal creature. Like, where is it showing me the agent.

And it's like, okay, just because it can be in any of those interfaces doesn't mean that it doesn't also have like, a physical presence. Yeah. And I think that's a much easier way to think about it, is that it is still a piece of software sitting on a computer. And that seems easier to understand, but that is what it is.

Well, it's also a complete it is a slight paradigm shift, inasmuch as it does ultimately make things easier because you can go through one interface instead of if you relate this back to tabs, instead of 17 tabs open to do one thing, I can be in one place my harness of choice or even Slack, and then I can access all of these different pieces of software from one interface, like.

Absolutely. And now you don't have to think about what all those interfaces are. You're just talking to an agent, and the agent's job is to worry about all of those interfaces are and to be able to bring together that information is the way that I experience it too, is I don't need to download something from one tab. Tell it to take that in some copy and paste it yet.

So copy and pasting is lower down the list than it is. Absolutely. Let's actually go through an example of what that might be able to do. We've chosen one here of social listening. Yeah. This is like a basic process in this traditional workflow. You might have like a CMO that spots that we're not answering Reddit complaints or reviews or whatever it might be.

Marketing lead picks that up, sends a task to an analyst. The download the or go and review the latest comments. Assess them drafts and replies. Send it off to legal I don't know, like it's different in every company, but let's just say it's 11 steps, seven people, seven context handovers, sometimes 19 days, sometimes just don't reply at all.

In an agent flow, we can significantly reduce the number of steps and the number of people involved in context handovers and the actual time to execution. That might mean that we bother doing it, which is same initiation. CMO notices that we're not replying, starts a task in Notion to the workflow owner, but then an agent picks it up, follows the brand guidelines, scrapes the relevant properties, draft a reply, pings that back to the workflow owner on Slack.

Like we were just saying where the agent lives. And then they say yes and it gets posted back via an API, and we're now live with the post. Of course, that could also just be fully autonomous, where the whole thing happens behind the scenes. And this depending on the level of legal risk or requirement within the business, like you might actually decide someone does need to approve that and you put that back in and it's not fully autonomous.

It's one human in the loop, which I think is just one way of articulating and a very basic example how this can be used to speed things up with the help of agents and just deliver better, higher quality, more consistent word. The interesting part that we were discussing earlier is that, like the implication here is that like, oh, there's only one person.

I hope you're enjoying this video. Every like, subscribe or comment adds to Danbot's training data and does its own part in helping Danbot take over the world. Thank you. Now back to the video.

Yeah, absolutely. That's right. And so here you can see the different people involved in building these agents. So an agent doesn't build itself. It needs humans there. It needs hundreds, maybe thousands of hours of experience of that person, that specialist in that area telling the agent what to do. So we're distilling that all into a skill or an agent to make it work and make it do exactly what we need to do that step.

Absolutely right. And then even like just to take that from one example, I mean, every company is going to be different for that. But let's just say that we said, like the legal team have had some input into what's approval and what's not, what needs to go through guardrails. That's just applied expertise that's calibrated over multiple examples. It was just how these agents can get built over time.

But of course they don't. As you said, they don't build themselves. And it's not just about people can't randomly pick this up. No, you need you need that expertise to help build out what's what's the intelligence. How do you iterate on that? How do you make that agent better? And we've said before, like you don't want a 100 IQ agent, you want 150 IQ agent, 180 or 180.

And that just comes from from building it and years of experience for people who have been in that area and then trained that agent and then what sage and understood how it should work and how to make it better. There's another layer to this which we'll talk us through, which is actually it's even not just with those people.

They can't just randomly decide, oh, I'm going to train an agent. You need some engineering in the middle, which is a team that. Absolutely. Yeah. You need the engineers there still, you need people who can understand how to make these agents work, how to understand the evaluations for those agents. What does that mean? What are you looking for?

How do you iterate and make these agents better and better? And you need a whole engineering team there to build out that infrastructure. So there's a lot of skill needed to actually implement this as well. And like choose between different models, for example, like it happened in the back end, surface it back to people when it's not working so they can can interact with it in a normal way rather than like writing code themselves.

Yeah, because it's not just about, oh, how do we make the the agent the most intelligent is how do we do it efficiently as well? How do you choose which model is going to give you the best result? And looking at all the metrics that might show you this is the most efficient model to give you the best output.

Rock and roll. Let's actually watch this run live, which is working. This example through request comes in to respond to customers on Reddit. This is obviously not correct. It's going to pull some juggling mucus. No one complains about us on Reddit yet. In this example, it pulls it up into a Notion table, or you can pull it into a Slack review process the workflow and it goes back in reviews, the reviews.

It's got three different suggestions. In this example it could just be one. They choose one, they press it, and then the agent will go back and post this on the Reddit page using the API connection. I mean, it could also, I think that you can build a custom web app if you need to elevate this internally through multiple levels like one approval layer or two approval layers and send it to different people, which is much easier than it used to be to build custom web apps.

And. Yeah, absolutely. So that's it really like it can have an approval process. It can happen autonomously. Just one very, very basic example of a one workflow amongst the thousands that you can imagine in many companies and the hundreds and that we have within our marketing agency. So this is just a way to make it happen more reliably, combining the years of intelligence of multiple people.

So more consistently I suppose, as well. And yet spurs in all the different areas so that like someone can apply judgment. And again, it's not about losing people, it's about distilling that expertise and essentially allowing you to make that happen more scalable, which I think is awesome and why we love it's workflows, not they don't work out the box, but when you invest in them properly, they can be really life changing.

Thank you for tuning in to learn about workflows. We will probably do a follow up episode, going even deeper on how to build these for yourself. But just to give you a taste of that, now you understand how to build flows.

Out.