Hello, and welcome to this podcast, Agents of Data.
My name's Joe. I'm one of the leaders of our agentic AI platform here at Matillion. And I'm joined today by our co-host, Julian Wiffen, our chief of AI here at Matillion, and our lead software engineer, Sam Perrin, who has designed and built Maya, our agentic AI platform that sits inside of the Data Productivity Cloud.
In this episode today, we're gonna get into what does agentic AI mean? What are some of those additional artifacts that are new to organizations that they need to be thinking about? And also, what does agent-to-agent communication look like in the future? And what does MCP stand for?
As well as other things like understanding the importance of context and how businesses can start to leverage the context inside their own organization to build on their agentic AI strategic initiatives.
Well, welcome, Sam. Lovely to have you with us, Julian. Fantastic. How are you?
Yeah. It's good to be here, mate. Good to be here. Doing well.
We'll do some introductions then, shall we? Get us all acquainted.
So, you know, my name is Joe. I'm a Go-to-Market AI specialist here at Matillion. I've been working with both of you for quite some time, and this is this is our idea that spun out of San Francisco where we launched our agentic AI product, Maia, about actually how do we talk about what agentic AI is and also how, like, you know, end users understand that. And we thought, perhaps, you know, this podcast would be a really useful position to help others understand a bit about what agentic AI means, but then also listen to some of the, some of the wizardry that is behind the scenes. And so that's why we've got the two of you here today. The wizard on my right, Julian, would you like to give an introduction? Yep.
My name is Julian. I'm Matillion's Chief of AI, and my team tend to lead the research and investigation into how we can bring AI capabilities into our product. And we work hand in glove with the man on my right.
Yep. Thank you. I'm Sam Perrin. I am one of the lead engineers of Maia, our virtual team of data engineers. Also a big advocate of how we use AI in engineering, coding agents, those sorts of things. So, yeah, just getting involved in all that. It's certainly an exciting time, isn't it?
It is. It's an exciting time. And I think the format of this conversation, this podcast is gonna be around how do we actually tap into both of your expertise in designing and building agentic AI systems and also trying to understand what is an agentic AI. What does that mean? And, I guess that's our sort of first question for today: Do either of you wanna take a stab at, defining what agentic AI is?
I mean, it's quite hard to define, like, people define it in different ways.
Mhmm.
At the end of the day for me, it's about using AI to power complicated processes, but you use the AI as the kind of decision-maker in there. It's not a human, it's virtual, autonomous or semi autonomous in some cases.
The analogy I like using is it's like moving from a single AI LLM process, which is like an individual contributor. An agentic system is more like a manager with a team of individual contributors; a team that the lead agent chooses which members of the team it wants to play, which tools it wants to invoke, and each one has spent more specialized capabilities.
Anthropic have a really good webpage/blog post they did on this called, ‘Building effective agents that details the different kind of agents you can build from workflow orchestration all the way to, like, fully autonomous decision-making agents that call different tools and things like that.
I highly recommend people read that, to be honest.
And so we're listening to both there, so we're just trying to distill and describe what that agenticness is around AI. It sounds like it's more than just another large language model integration that you can use for a single purpose.
It's actually about empowering something, some agent that's in control of other agents as part of a framework, and that's what allows you to have the agent specialized in different ways.
Like right. There might be some tasks where there's large volumes of text to process, but the the level of reasoning you need is not very high. So you might use a simplest, smaller model for that that's faster and cheaper. Okay. And there are some when you need a higher level of reasoning. You might bring a different model in that space. Others are other tasks might be ones where you wanna get very specific, wordings to the prompts to make sure that they just got a very specialized remit to, like, only do one thing.
But you're describing, like, different roles that are currently taking place in, like, businesses and organizations today. Like reading it through reports or reading through PDFs or transcribing calls. And you're describing these as things that agentic AI can do and help with. But then underneath that umbrella term of agentic AI, there are these tasks or there are these particular processes that businesses and our customers today are already aware of. And, do we wanna talk a little bit about, like, what's in scope? Yeah. Like, those types of tasks.
What's not in scope at the end of the day? Right. You know, most information technology-style workflow processes, you know, and me as a software engineer, agentic coding is the thing at the moment. And then we're in the data engineering world as a business, and feel like data engineering lags behind software engineering, but we're, you know, we're getting towards that. And now you've got entire engineering teams becoming 10x more efficient by building and using coding agents such as Cursor, Amp, Claude Code is my personal favorite at the moment.
And these things are, you know, that's just coding. You've got data engineering. In our case with Maia, where we're building data pipelines through the natural language, through the agents. But then all the way down to you know, you're doing deep research in Chat GPT. You're using that as a ‘planning your itinerary for a trip to Seville’, for example, in my case. Like, that's an agentic process of crawling the web and going and picking information off of there.
So, are people already using it? Like, that's a really great point. The last one, if you double click in on that, are people already using agentic AI when they're interacting with some of those LLMs in that UX interface. Like they're just unaware that that is it?
You've got a system instead of you might be talking to one, but in the background, that's talking to other LLMs that have got slightly different remits.
Okay. And or invoking tools because the tool invocations are really just a set of instructions that say, ‘Okay, you can call this tool to get a report or or to search the web’ and get a doing the equivalent with in Sam's Seville example, you might have one with a tool called web searching. Like, okay. Go and Google hotels in Seville and bring back the results of that Google search to the main agent who then says, here's the three hotels I think match what you've said you're looking for best.
The key part in that workflow is the agent part in the middle. It's not just a prescribed process of go and find me ten of these things, are they relevant? It's the agent or the LLM in the middle there deciding whether it thinks they're relevant.
Or maybe I've not got enough information yet, so I should go fetch some more information.
So that agent's choosing what to Google, in fact. Yeah.
Yeah. Yeah. So, yeah, you were using some interesting terminology here. Right? Because, like, we're saying that this to be an agentic AI system, there is something in the middle that is deciding, that is making a decision based on the information that it's being fed.
And that's the kind of the biggest league we've seen, right, in the past twelve months of going from just that single interface with LLMs to now actually having all of these communicate together.
Some of it's about also constraining the output options that it has because it's decided to use.
So previously, with your general ChatGPT interactions, you put some text in, it gives you a big block of text and answer. Whereas, if you start to put some constraints to say, okay, here's the four different levers you can pull. Here's the four different answers you can choose from about, okay, respond to the user with text, go and Google something, look at some other source of data, call another tool, or call a tool for the booking for a booking line for one of the hotel systems to actually place the booking.
Yeah.
And you're forcing it to say, okay, you've actually gotta pick one of those four options and decide what you're gonna do rather than just give me unconstrained text.
Okay.
Wow.
I mean, so that offers a lot of, you know, impact for where it can be used. Maybe you're like I think some of our listeners, some of our customers are already thinking about, well, yeah, but what does this actually mean for, like, where I could deploy it inside of an environment? So, like, perhaps we'll come on to, and you mentioned Maia in your response there, Sam. We'll come on to that in a little bit about what that offer is. But just taking a step back out and thinking, if I'm a, you know, a chief data officer of an organization that is, you know, day-to-day being fed with all of these different AI products that can offer similar levels of agentic AI, but also actually taking a holistic view of where do I pick which things that we want to have an agentic AI system play. Where do we see, like, that going in the market? Where do we see that sort of funnel of discovery going down to, well, this is where a really suitable role for an agentic AI system is?
It's quite a hard one, so I thought I think people are trying to figure it out still, but the the most obvious ones, to start with seem to be the ones we had simple tasks.
We used to talk about the infinite intern, which was the ‘if I could give somebody some simple instructions and go and sit in a room and do this repeated task, it would they didn't need to know anything else’, but ‘I want you to go and read this document and process it or flag it or give me some sort of information out of it’.
That was a suitable task for an outlet.
Now it's their scope and their skill is expanding significantly. And we're looking at the places where you still need a lot of human thought, but the agentic framework can grant maybe a junior member of team skills of senior or allow a senior member of the team to really scale up. So the honest obvious place is the way you've got really resource constrained bottlenecks. And that's where we see in data engineering because there's no data engineering team in the world that's got excess or surplus capacity. They're always bombarded with a request for more data or more analysis.
And something that allows their teams to scale up is a very powerful therefore, a very powerful offering.
As I said it's almost identifying the human-workforce bottlenecks that exist in that organization today being then a primary driver for where you think about deploying agentic AI.
And that doesn't necessarily just need to be data engineering. We might be coming to that in a second. But, Sam, how would you approach it if you were thinking because you do a lot of development in agentic AI systems already as part of your, like, day-to-day role. But how do you personally, like, prioritize where you think you know when to use agentic AI and when not to?
Yeah. It totally depends on the use case. Like, I mean, often, like, of obviously LLMs which power all this, it's text. Right?
They use text or voice or, you know, if you have voice data it gets turned into text and then fed to an LLM. So anything textual code, data, like most of the format of data we have, we have CSV files everywhere. You just have even your instructions or, you know, another use case that I might use day to day is like no one in software engineering likes to write Jira tickets or epics and things like that. But at the end of the day, I can now take a few bullet points of text and turn that into a complete completely fine epic of user stories. Flesh out things out or anything.
The fact that we can now do with this stuff that previously you couldn't do programmatically because you need to precisely codify every step or every decision. Okay. But now you can give a set of bullet point instructions to say, okay, check this against our standards, check this doesn't break any rules. And only in your nonfunctional requirements and other sanity checks.
It's still unstructured data, isn't it? Like, I mean, it can work with structured data, but, like, that's generally already solved problem. And obviously you can have semi-structured data, but unstructured data, this is the world we're in now, like previous use cases of data engineering for example, where you just have like call transcripts, like bucket loads of them. Well they're just like literally people's words in like a massive long transcript. Like previously it would have been really hard to pick that apart and analyze it, and now you're literally just picking off the shelves model like OpenAI GPT or Claude, and suddenly you can just drive all these insights out of it. That's like untapped knowledge. Yes.
So by targeting simple, asking for simple judgment calls on that unstructured data.
Yeah.
And you're using ‘are they talking about this topic in the first part of the call?’ That kind of Yeah.
And it's the combination of now you've got like software engineering problems where you've got this audio file of like a call, and now you use like a managed service to extract the text, the transcript out, and then you file the transcript into an LLM, and it's just like, wow, this is a day I never thought I could extract an hour.
Unless you combine unstructured and structured sources as well.
That's oh, oh, let's let's double click into that.
Yeah.
So, so what I feel like we're gonna edge into data engineering territory now, but which, you know, to keep it high level. So you might, the sort of things where previously, you might have a human who would do a lot of crunching reports, but I have to read a whole lot of text and then summarize it. Yeah. And then I might look at a bunch of numerical data alongside that regardless of whatever I'm doing. Whereas here, you could have the LLM do a lot of that work. You know, it might be, let's say, I'm looking at customer churn prediction.
Yeah.
The data scientist world is very numerical models looking at how many support cases they raise, how much they're using, what price they're paying, all those kind of things that feed into an old school numerical model. But here, we can put that alongside here's all the call transcripts we have with them. Here's all the support call notes. Here's all the account manager's notes to say that the customer is really unhappy about x y z. And use that to to get produce a a summary that's sort of humanly consumable scale of, like, a paragraph or two to say, here's why we think this one's at a high risk of churning, and here's the two or three bullets out of this fifty pages of data that are the mode that ‘here's what they're really unhappy about and what the heart of the issue is’.
And and going back to, like, that agentic way of thinking about solving that problem, like, there's there what is there a risk here for a CFO or or or for a decision maker in a business making decisions based off this data that, oh, actually, the AI is almost prioritizing for me what a combination of that looks like. Like, you know, let's unpack, like, what that risk means for a business because, like, there's, you know, there's guardrails that there may need to be put in place, but, like, to help our listeners understand what does that decision process look like today?
Like, do you I mean, I don't know who wants to fill that one off first.
I mean, like, do you mean like the if I was gonna use or let's say, alright. Today, build an agentic system, how do I know that it's actually performing and answering in the right way?
Well, how would I, like, how would I be sure that the quality of the because the if the agentic AI system is making a decision on the free text summary that's coming in from the goal from the calls, but then also the structured data that we have in our database, and it's making a decision based off of the blend of that.
How do we know what the quality of that decision looks like?
Yeah. So there's I mean, there's a few ways. Right? The basic one that probably everyone will still do for a while is human in the loop.
Right? Like Right. These processes might result in an action or something being generated that effectively, like, you know, pays an invoice, but, you know, you still want a human to go, here's the invoice, here's the data that built this invoice, I'm gonna hit approve on that manually, and then that goes through and gets paid or something. Okay.
Then the other side of it is as you start to build sophisticated agents - Julian and I experienced this right now, it’s evaluations, of the model and its performance over tasks, building a framework around that.
But it it's really complicated, really hard to do.
You get, you probably go in the scale, kind of scale path like Sam said. You start with you in the loop. Yeah. Alright.
But it's about building up trust and confidence in these things and or understanding whether it's doing the right thing. They that kind of decision making would be subject to quality control in no different way if you had a bunch of humans doing it, really. That you're like, okay. Are my juniors making the right calls as they're reading through the stuff and summarizing it ridiculously well? Periodically, you're gonna say, well, show me where you got that information from. Right. Build confidence in that person that they've got us out.
They're judging things the way you want. So with the agentic systems, what you're probably doing to start with, obviously, as Sam said, you start with the human in the loop. Like, okay, it's doing the prep work, but a human's always hitting the button, it's always reviewing it.
Okay.
The next step is probably then, you may be only the human only reviews a small sample. You know, your quality control, like a factory or a process. You might say, alright. We'll inspect five percent of the transactions and see what and check that the expert still agrees with the large language model output.
You can also then start to look at AI as a judge where you use another large language model to review the decision and say, does this comply with these criteria? Does this meet our standards?
But you still wanna calibrate that with a human, but you then, you get into the business of, well, do an exercise of marking the judge's work to say, is it is it scoring correctly? Do I trust the judge? I can now trust the judge. I can scale up hugely.
Do you think we're getting to the point where it's where now we're gonna see agentic-to-agentic cross checking and references to the judges.
I know.
It's clear already. Asking a second opinion.
We've all like it in our brain model process for Maia right now, in the behind the scenes. We actually have a judge. We actually have a judge LLM. An LLM as a judge that is the user and drives the conversation because it's non deterministic. So you give that user a role and a goal, and it has the conversation with Maia. And then on the other side of it, we're actually building an evaluation agent that looks at the result of what that conversation's had and evaluates the conversation in the unique ways that it may have played out.
Yeah. So so, let's okay.
Let's dive into that in some more detail. But first, should we set the scene a little bit on what Maia is? Because we've used the phrase a couple of times. For us here at Matillion, it's become our, you know, biggest strategic objective and goal that we're working towards building into our into our platform, the Data Productivity Cloud. And the Data Productivity Cloud, for those who haven't used it before, is that easy to use interface that allows to move data from your on premise cloud hosted systems into your Snowflake, Redshift, Databricks, and it's doing all of that in a YAML based file, which is visualized in a drag-and-drop interface.
We've got leading CICD within the product as well.
And now we're bringing Maia into that Data Productivity Cloud space, our agentic AI system that is helping end users be more productive with their day-to-day tasks. And so that's kind of a little bit about where Maia is currently in the Data Productivity Cloud. But you just touched on a really interesting point that I wanna come back to, Sam, which is there's this ability within Maia today for it to act as a judge and to actually be driving that conversation interaction with the human, but setting out the goal that the human is trying to achieve and and almost what? Self correcting on, as it goes through?
So it's a little bit different than that. It's the, we've effectively so we have Maia. Maia's the all glamorous Maia, which you as a human converse with. We then build other agents that aren't technically Maia that simulate a human user using Maia to drive a conversation to achieve a goal. And then we use a third agent to evaluate and score that conversation.
How successful was the pipeline? Did it do the right things and that sort of stuff?
And that lets us test at scale because obviously with a non-deterministic system like any of these identity things, you got a randomness.
So anytime you wanna do tests, it's not like a single set of unit tests. You wanna run them ten, twenty, fifty times Mhmm. And say, okay, ninety three percent of the time it got it right.
Is that good enough kind of scale? Or twenty percent of time go right, no. Clearly, we need to fix that use case. But you can't you can't get that level of scale with with human resource. You've gotta have a way of automating it. And this this is approach is how we do that.
Okay. I mean, this is brand new news to me. This is sounds really exciting. There's agentic systems that we've built, Maia, that's also talking to other agentic systems that are performing scalable tests, almost simulation. Yeah. Environments that we can then increase the overall intelligence of Maia.
Yeah. So it would let us test, like, variations of how we set it up, what models we use, and the like at scale.
Okay. And before we get into the detail of what Maia can offer, like, as in how it can be adapted to a particular environment, if we could take a step into the world of, like, more just agent-to-agent communications, for a second. This will be, you might have heard, like, A2A to a or MCP.
This will be quite new to a lot of our audience, like Yeah. We should be just scene set for a second, and what does that mean? Why is it important?
Yeah. So you summarized two there, and they're like the common protocols. There's a few others, but MCP, kind of, came out of Anthropic originally, but is probably the most adopted one now. Now LLMs, the good LLMs or the more recent ones have something called tool calling. That's where they, they generate like a structured response that you can then programmatically call within your code, and then you return that response to the LLM, and the LLM kind of reacts to it.
Those tools you often define as software just as you would if you're writing any software. Right. And what MCP does is it's like, discovery, like protocol, a bit like rest APIs were for kind of standard interactions of, like, microservices or just software.
MCP is like they're calling it like the USB port for, for AI and agentic systems.
So USB C. Yeah? Yeah.
I hope, I hope so.
Alright. It's really just the manual or brochure for the agent to read. So it the same way human use would read it and say, okay. Here's how I call this API to get information.
Telling the agent how to do it. And the agent's going, ‘sounds like what I need’. Alright.
I'll send that call. I'll configure my request and send that to get information.
So this is peeking into the future of, you know, of how these systems are gonna communicate with each other.
Yeah. Yeah. I know that. My voice is very croaky.
The A2A spec is like a whole level above that where whereas we have MCP that will kind of declare the tools that are available for a service, and then that might be something like, at last year and release like one that prepares tools how to read Confluence files or create Jira tickets. A2A is quite different. A2A is maybe not as adopted yet. It's a bit more new and that there's not a lot of agent-to-agent use cases out there I don't think.
But that's a similar protocol, but all based on like tasks and execution. So an agent will say, I'm this agent and my task is to do x or this is my role, declare itself using like a well-known agent JSON file, And then your agent can look for those agents and just call them and execute the tasks that that agent has. And it can ask for the status, and it can stream it back. And it's all pretty cool.
That sounds amazing. So this is this is so where A2A is essentially greeting another human or is that access control request that comes in. Do I have permission to do this? And here's what I need to do. And it's giving them the here's what I need to do to that other system. Okay.
I think I think I'm getting it.
And we're seeing that could be a view of the future of what we need to serve with Maia, which is, at the moment, obviously, we're gearing Maia up for human users. But we expect it to be other agentic systems talking to Maia, particularly because we're in the data engineering space. And they're wanting they're gonna want to request data. You might have one from a data visualization platform saying, ‘Hey, Maia, can you bring in this additional table and add these columns so that I can bring this into my report?’ Right. And the the data visualization tool might go, I haven't got that. It's gotta pass that request on to Maia, grabs it. Maia works on how to bring in the new data. Yeah. Reports back when it's done, and then the visualization system can pick it up and say, okay, now I can add the extra columns to the chart or whatever it is.
And so that but that end-to-end process that you were describing there, that's taking us a series of teams of human people today in all of our customers based, in all of our organizations globally, a lot of time to get that visualization built or get the answer that they're looking for all the way back from those source systems and marry that up. And for those of you, for those of us that are not in the the world of data engineering, that is kind of what data engineering is all about. Right? It's connecting those source systems to those target systems to help get that visualization downstream.
So double-clicking into what does data engineering mean for this a to a communication in the future, are you saying that there's gonna be a space where Maia is this a to a communicator with other systems?
Is that yeah?
It could be either way. Right? I mean, it could be that you have a BI dashboard, and you ask it a question, and that BI tool realizes it doesn't have the data it needs. So then calls Maia to build that data, pull into the, you know, the bronze layer, do some sort of refinement of that data and then leaves it there and returns back to the BI tool and says, ‘Hey, this data is ready for you’.
And then the BI tool goes and does its SQL generation and builds a pipeline. But on the flip side, we might eventually have, you know, other agents that we wanna talk to maybe, Atlassian again, I'll use them as an example. They build an agent that can just like look across Atlassian, answer business questions in your Confluence or Jira, what would be really useful for a user of like Maia for example, would be if we call to their agent and say, ‘Hey, users are looking for this piece of information, can you find it?’ We talk to their agent and we get it back.
Yeah. So, amazing. All sorts of opportunity there. But it's definitely like MCP is by far the big, the biggest one at the moment.
Right.
Okay. It's a little bit easier.
And what forgive me, what is MCP? And what is and what is that? Is that something something protocol, I'm guessing?
Model context protocol, I think.
Okay. So let's talk about context then. Like, what's that, is that relevant to building a agentic AI system? What is the relevance there for that?
Yeah. And we're passing the right context with the question. It's always important for any Q and IPs where you're giving the background. Okay. What's these terms mean? What do I want you to think about?
And a lot of it is just with the MCP side is how one system frames a request to another and how do I format it?
So I'm booking my holiday to the beach. Yeah. And I'm gonna be giving it the context in Chat GPT or in whatever large language model that I'm choosing to use, Gemini. Yeah. So, so like I say so my context bit is is what? Is it me describing what I'd like to share?
It is everything. Or is it I guess everything. I know. And this is the, like the reality is your prompt is a context, and then the tools are used to do retrieval augmented generation (RAG), which is a fancy way of saying go and grab relevant content and include it in the prompt.
Tools are used to retrieve information or potentially, like, I don't know, manipulate state in a system somewhere. But a lot of the time, booking your trip to the beach or whatever, it might go and use a web search tool and that web search tool is, you know, model context protocol says here's a web search tool. This is its input. This is what it returns. So you query it, pull in the context, and then that context is then in the prompt.
Yep. And, like, the context for the web search tool might be, okay, what's the search term you're going for?
How many results do you do you want me to give you. That kind of level of, but this is a whole new that's just a whole new way of thinking about in terms of in day-to-day business, about how do we, like, provide that right level of context to our end users that are maybe wanting to build agentic AI system?
But then also sometimes you're formalizing what you in a human process would be an informal agreement. Everyone, right, commonly understands that if I request something from another team, say I request a report from another team, I probably need to explain what data I want. I need to explain when I want it by. I need to say what format, what what platform I want the output in or what format I want the output in, those kind of bits of information.
And it's formalizing just enough of that the LLMs can provide it, even though it doesn't have to be a programmatic, every variable perfectly filled in. Okay. But it's just giving the right shape of what's needed.
And so that's a whole new field of view and place for businesses to start thinking about preparing for the the creation of the context and and where that context sort of sits with inside of an organization and how and and how we then set up the MCP to be more effective.
Where do we see, like, that going?
There's probably a lot of complexity in managing it and checking the quality.
Right. And just you've got you built up stuff that's potentially there's a very complex audit trail. So let's say this makes a mistake. Yeah. And there's been three different agentic systems talking to one another. You need to be able to track back which, what, where those errors and hallucinations have crept in.
So I book my flight or I say that I'm going to the beach. And then the first communication is to a flight checker and it books it through an airline. And then there's a mistake made between the airline and the my calendar that's talking to it as well. Yeah. And what you're saying is we need to provide the context for which parts of that process are auditable, where is it trackable. We've you know, it can't be black box.
When you turn up at the airport and they say, ‘we haven't got a Joe Herbert on this flight’, somebody needs to know why that what the hell went wrong.
Okay. And then how that's addressed. And that it you know, it's no different from any other data processing protocol. But, like, you've got multiple complex systems.
How do you manage them in a way you can get reliability?
Yeah. That I mean, that is just like a software engineering problem though. Like, the tools tend to call APIs somewhere in a software system that's been built, and those software systems will have, like, auditing capabilities and stuff all built in with it. So Yeah. It's just standard process.
Some of that I think is tracking that you get consistent results each time sort of thing. Like, you query the holiday on Friday and get these results. You query it again on Saturday, and it's just completely different places, completely different prices. Yeah.
Is that correct? Or is something, has something changed there that it has misinterpreted?
Yeah.
And, okay. So there's, so, it's still that context with inside of that agentic AI system that needs to be tracked to make sure that it's doing what it's supposed to, giving consistent answers, it's giving the answers you want.
So that and the holiday firm doesn't want their agent to offer you a flight for ten dollars. Gotcha.
That kinda.
Okay. So how do we scale that concept up when it becomes in the field of view of data engineering?
Like, what is what and and, you know, I'm leading into what we can do with Maia, yeah, already today.
But just what are the kind of that outside parameters that we as data engineering leaders need to be thinking out when it comes to designing how to ingest this concept?
It's how you get repeatability.
Right. And that plays to Maia's strength, which is, let's say another agent existing calls Maia and ask for information. Okay. Maia's gonna build a pipeline to get that information. Okay. So the next time somebody asks, it can refer to that pipeline and see it's already got it rather than building it again from whole flesh and maybe putting business different business.
So it's so it's thinking intelligently about using its assets that it's already got available to it as opposed to just creating net new. Yeah.
And a and a key point, I guess, in data engineering, one of the big artifacts or whatever we create is SQL. Right? Like, just generating a big block of SQL and then execute it somewhere doesn't really cut the mustard, you can't validate, you can't test it, you can't step through it. Yeah and I mean it's where, you know with Maia and we kind of have that process of building a more of an abstraction on top that goes through a lot of processes and rigor around making sure that the pipeline we've built does the right thing and having the the mechanisms to check as we build the pipelines that's good.
Yeah. It's data coming through at each step as it means. Yeah.
And so we as human data engineering leaders today need to be taking a step up into this world of agenticism to understand what the provision of those types of files and rules actually looks like. Like, when you were getting started, like, thinking, you know, developer mode on now, like, when you were getting started with playing around with some of these other tools that you're using, like, how did you do that discovery, that self learning? What did that look like?
As in self discovery, like how to manage the context and things like that?
No. More like like, for your own personal, like, development. Like, how did you how did you start on that journey of, like, looking for, like, the right sort of guidance and noise from market, from your peers? Like, you know, you've, yeah, you've built You've been building here, you know, code of Matillion for, what ten years?
Not far off. Not far off. So you're already well adapted into, like, how to build and write really scalable, repeatable programs. Right? But then there's always, like, there's some I know I know we talk about this all the time, like, you're always trying to, like, think about what's next, how could that be better. What advice would you give peers that are thinking around, like, where do I go to learn how to learn and upskill here?
Always by doing. Like, I'm self taught software engineer, so, like, everything I know, I learn by doing. So you can grab these tools, you can, you know, you can pick a change up, you know, we use Jira to track issues. Well go grab a ticket, go have a go, were you successful, were you not? Like maybe try and change the prompt, just just work with it. There's lots of resources out there for people, you know, there's varying degreeing ends of the spectrum now where you've got people using things like Claude Claude code and spinning up like twenty different terminals to spin up multiple engineering tasks, and then you've got the guy that's just still using co pilot like it's chat GPT or something, and there's you know, we've got a everyone's gotta get with the agentic way of working.
Is there oh, the way oh, the agentic way of working. Let's come into that.
If I could actually say it, that would be good for the other day. There's someone losing their voice.
I think there's a strong trend that I feel, I mean, it's brilliant in the same buckets now. Folks that are that folks that tend to apply the learn by doing approach seem to do well in this space or seem to be I think a lot of us that have ended up in this area in that. But because it's moving so fast, if you wait for somebody to build you a training course, it's almost out of date by the time you watch it.
A hundred percent. Yeah. Yeah. And it yeah. Absolutely. Like, and then there's this constant self updating context, like, as a human trying to think about these problems of, like, well, it couldn't do this when I built this six months ago, but now what if I try it again now? And, oh, that's actually hold on.
We've done a whole bunch of, ‘this is kinda working. It's not, results aren't quite good enough yet. Alright. We'll park it for three months or six months and come back to it. Don't bin it. Just think save it. Save it where it really is’.
And the advances in the model to and and models and other techniques often then unblock get you from almost working over the line to this is now good enough for use.
Yeah. We've seen, I've seen so much of that. It starts with doing. Right? It starts with just You're learning while you're getting.
And, you know, we were learning with those janky chat bots that impersonate and made your boss talk in a silly voice or whatever it was. Yeah. But we were learning the techniques as we played around with it.
Like, the important thing if I can actually speak is not not because I'm interrupting. Sorry. I really is it because I can't get my words out today, is that, like, you have these advocates and, people in engineering, like we do. And there's like I said there's like the two ends of the spectrum where, the real experienced ones of agentic coding tools become your advocates for all the people in your business that aren't really there yet. And that's not because they're doing something wrong. They might just not be the person that gets to grips with the coding tools because it's a completely different way of working.
The plus of the code the the coding tools or the assistance is it lets you experiment faster so you can learn by doing more.
And then I'll also go back to Maia. Like, I've had it recently where it's where I've asked Maia, how can I give you better prompts? Yeah. And it's actually intelligently aware enough to then to then generate more information for me about how I can change the context files that I've put into my project and the data productivity.
They're probably worth a discussion actually, the notion of the context files.
Yeah. Absolutely. Yeah. Like, let's set the scene here. Let's see where we are, cutting edge delivering of Maia and agentic AI platform for our customers today.
We're leading the way in what the future of data engineering looks like. And a big part of how Maia plays a role in Matillion's future is being able to adapt to the context in the data landscape of those customer environments. And so to help Maia then be more performant in that space, we've got these context files. Now, what what the heck is a markdown file, Sam?
Like, just like, I've I know that they're cool, but I still don't understand what they're, what they mean and what they play a role in.
Yeah. I mean it's just like a nice way of defining documentation, but it's like code. It's it's richer text basically. And what you'll end up with is like, I mean, people who use GitHub or any markdown renderer, it just you can use special syntax to render the document more nicely. So you can bold headers, you can do formatting and highlighting.
And the reason why I think you tend to see context files in, in agentic tools using markdown, is because they've been trained on so much text and there's so much markdown out there, and certain things like, you know, a well-written document by a human being is better understood by another human being, because they've been trained on so much text from humans, they're actually really good and understand, you know, if you put emphasis around something because you bold it or you you have a section that's got a really, you know, a massive header above it and it's like this is really important.
Then they'll will “listen” to that in this document.
So, like, we like, I use bold and and upper case when I'm like, may please pay attention to this. Yeah. Yeah. This and then and then we're saying that actually the LLMs themselves can understand the context behind the I don't know what to use the context there.
And they can understand the importance and significance that a human is trying to place on these instructions by the way in which the markdown. Okay. Gotcha. Right.
So now we've got our markdown files. What does that mean for our customers in terms of how they can think about giving better instructions to Maia?
Well, I mean, it could it depends on, like, the business context that they might have. So, like, relevant information to the problem or the project they're in can be really helpful. Like, in we do the same thing in coding tools so we might include an extract about what this particular software project is for, because we have multiple of them. And then also like standards and naming conventions and code style formats and potentially how we run tests and things like that. It's the same in a data engineering project. Like, you add all the context you think is most relevant that that your coding agent needs to know. Right.
It could be stuff that's as simple as the raw data coming in this schema. Here's the URLs or connection details for a couple of sources you might want to connect to.
Little notes like saying anything any table that begins with ‘JW_’ is just Julian's temporary tables. Ignore them.
Gotcha. And, you know, 28-day CAR is the is the one of the ten different revenue fields we've got when somebody asks you how much money we're making. Those kind of bits that you'd learn by asking around if you're the human data engineer.
Gotcha.
So it's just providing their knowledge of the, it's like you bring a new employee to the team.
Right? Just go give Maia the basic notes, like, here's where you get started. Yeah. Here's what you're in there.
And I feel like that tees up for a really great conversation we might have next time on semantic layer and that importance of what that semantic layer plays like and looks like in the field. But, Julian and Sam, that is all we have time for today. Thank you so much for joining this session.
Thanks for listening to the Agents of Data podcast.
The podcast is brought to you by intelligent data integration platform, Matillion. To discover agentic data engineering, visit matillion.com.Maia. Don't forget to subscribe and let us know what you think of everything we discussed over on Matillion's LinkedIn and Instagram.