Microsoft’s Reasoning Layer Bet and Dataiku’s Agentic Pivot for Unified Data Intelligence

Control. This is Agent Dev. I'm in position.

Copy that, Dev. Stand by for go. Standing by.

Hi, everybody. Welcome. You are part of another episode of Agents of Dev Podcast.

My name is Mitch Ashley. I lead the software lifecycle engineering practice at Futurum Research, and of course, my co-host, Brad Shimmin. Brad?

Hey there, Mitch. How you doing? Good.

You're lead-- What are we calling data intelligence engineering fabric? SQL-ish? What's the name of your practice now?

Yeah. So- Things are happening so fast we can't keep up ... I'm sticking with data intelligence, analytics, and infrastructure.

I'm trying to cover- There we go ... every single base imaginable. There's a lot of territory there, for sure.

Yeah. You don't need me to tell you. And you are out of town this week.

You're not just at the shopping mall picking up your Rolex watch. You're somewhere else. Where are you?

Yeah, we're recording this while I'm in Atlanta, visiting with Microsoft. And they're doing a little shindig. This is the third year that they've done their Fabcon, which is short for Microsoft Fabric Conference.

And this is the first year that they're conjoining that with the SQLCon. Hmm. And if that isn't a statement of intent or direction, I really don't know what is.

That is a bit of a leading yeah, leading the horse to water there, isn't it? Right. So, and I say that because, and I've written about this recently with some of the work IBM has been doing with Db2, and everyone who's over the age of 50 will know what I'm talking about.

These storied relational databases, and operational data, and the value thereof, is really coming to the forefront right now in terms of bringing reliable context to agentic systems. " Those kinds of decisions demand access, timely access to accurate data. And my friends, that is where relational operational data comes into play.

So it's a big deal. I don't blame Microsoft for launching SQLCon. It feels like things are going to flip, and data's going to go from this back end thing to a front end thing.

Oh, I love the way you said that. Like, that is the workflow. Yeah.

Right? Don't you think? Yeah.

Totally. Because it is what drives everything. " No, it's all built into this agentic process we're ultimately looking to build, right?

That's right. And it lives everywhere in that life cycle, and everywhere in that stack simultaneously. What's that film?

" Yeah. " Yes. Yeah.

That movie. That was a great movie, too. So now, time for the call-out, for our call-out segment.

And one of the things we wanted to call out, you did a great market report around pivoting to AI success becoming a platform for creating workflows. Well, that's probably an understatement. Talk to us about Dataiku and your report.

Yeah. Dataiku, I've been tracking them for a long time, and I really appreciate this company. So it's a French company, and they come at the market from that perspective, which is for anyone who appreciates fraternité, liberté, egalité.

Am I saying that right? I think I got that wrong. The Tricolore.

For us non-French speakers, it's good for me. Yes. Please forgive me.

It means a lot to me. I'm very tired, so I'm not sure it's right. But anyway, they, I think, have always come at the market they played in, which was the MLOps market for operationalizing your data scientists' work.

Mm-hmm. And they tackled very early on agentic tooling and AI as an independent entity capable of taking action on its own early. And they are kind of doubling down on that right now in that they're actually kind of pivoting their whole company to, instead of focusing on the basics of MLOps, to instead focus on success.

And this totally matches what our research has shown, which is that companies are no longer investing in capabilities. They're investing in outcomes. And that is where the money is going.

It's like starting at the end in mind. Duh. And so- It's even coming.

Yes. So they came up with this idea of a sort of decentralized platform, a control plane, if you will, which I know that you are the master of, and so you'll appreciate this. They came up with a control plane for managing distributed, decentralized agents, across disparate architectures and different technologies.

It's really nice. So anyone who hasn't looked into them or that, really should look into this because theyThey're all about how you build the agents, how you manage the agents, how you use that to create within the company a sort of what you would call a reasoning layer or reasoning engine that is- Mm-hmm ... sort of like, you know how we were building databases back in the day to sort of represent the knowledge of a company?

Well, now we're building- Yeah ... this agentic layer, this reasoning system as that representation of knowledge of a company. Well, and we talked about this on our last episode about the explosion of agent control planes and everyone kind of coming to the table with theirs.

Yep. So with a Dataiku, what is the lens of what they do as an MLOps company in influencing how they're interpreting what it means to have an agent control plane? Yeah.

Or do you see that they've kind of taken a more broader view of it? Have they kind of put the guardrails on and said, "This is what it is for our customers, our market," or are they expanding into something else? Every company is the sum of its parts and its history together.

You can never escape- Mm-hmm ... who you are, and so it colors who you become, just like Keynes. And their capability has always been in sort of combining people, process, and platform.

That's what we used to talk about with MLOps as the triumvirate- Yeah ... that you're trying to achieve. And they've taken that now to be people, orchestration, and governance as their binding principles, if you will, for what they're working on here.

And so it's an extension of who they have been that projects forward into this new agentic world that we all find ourselves in. Interesting. Yeah.

We all have our own take, and we are a product of our history, even when we reinvent ourselves. Even when, yeah. As companies or individuals.

Well, let's turn... You and I have been... It's conference season, so if we're not on the road, we have virtual backgrounds that look like we're not on the road, or we actually have backgrounds that show we're on the road Yes ...

at a conference. And you're just coming off of going to Microsoft Fabcon and SQL Con. Where do you think that conference is going to be next year?

How is it going to be different? Given the sheer scope of the announcements that they made this time around, I'm not quite sure there's anything left to achieve. They seriously covered some ground with this show.

My goodness. I'm going to post a write-up on it probably tomorrow. There'll be some links where everyone can read more deeply into some of the stuff they rolled out.

Like for example, they rolled out this thing called the Database Hub. Anyone who uses databases and uses the cloud will know that AWS has 13 databases to choose from, Microsoft as well has a cadre of databases. Same for Google.

Mm. And they all kind of exist separate from one another. And what Microsoft is doing, and it's all kind of predicated upon this foundation inside of Microsoft Fabric called OneLake, which is really a lakehouse, but they call it OneLake, and with Fabric is to sort of create this plane of managerial abstraction wherein whether it's SQL Server or Cosmos DB, or Azure Database or Postgres, name your database, and even databases that aren't running on Azure, databases that are running on- Yeah ...

premises. This is through this- Mm-hmm ... technology called Arc that they have.

But anyway, Database Hub sort of creates this control plane across all of them for, and I know you'll appreciate this, Mitch, for observability. Observability. I've heard about this before.

It is somewhat critical. So, anyway, that's just one announcement. They also introduced a new...

They've been working on this for a little bit separately, this thing called Fabric IQ that sits next to... They have Foundry IQ, which is for how you get data in, and they have... I forget the name, but apologies for the noise.

Somebody's dragging a cart through here. They have another IQ that's for Office 365, so what you do at work, where data comes from external to the Azure platform, and then the Azure platform. And they have this idea of a semantic layer, which is actually they take it a step further into what we call an ontology, which describes not just the meaning of things, but how you arrive at values from those things.

So when you say end of quarter revenue, well, how do you calculate that? What fields does that use? Where do they live?

Mm-hmm. What does that mean for different departments, et cetera? All of that is sort of spread across these IQs that Fabric has.

And with that, developers can build on top of any of these databases. And with a few of the tools that Microsoft has at the ready to support developers, because they do have a couple that are mission critical for most developers- Yes, I've heard that. that you can very quickly build agentic systems that haveAll of this semantic and ontological knowledge at the ready wherever the data lives, because it doesn't matter, because it all kind of gets what they call mirrored or shortcutted, that's a bad verb, into one lake, so that you have this very consistent view of all of your company data that is metadata enriched, so that when you write your code, you don't have to worry about, "Oh, gosh, do I have the right database?

Do I have access to the database? " It's brilliant. So that's why I say, I'm not quite sure what next year's going to bring, Matt, because they seem to have a lot of the big challenges that we're facing as an industry sort of under control or at least envisioned.

Envisioned is what I'm going to go with, as a matter of fact. Envisioned. Because a lot of what I'm talking about is actually in either private preview or public preview.

So it's not- Yeah ... all of this is in GA right now, but a lot of it is going into production right now, like being able to bring in data from SAP and from Salesforce and Snowflake and Databricks without moving said data. How does that sound?

Mm-hmm. Sounds really great. Because, well, the progression is, if folks haven't figured it out yet, is we go to preview, we go to GA, then we talk about customer stories or use case of case studies, right?

The first one is usually about three to six months apart. Yep. So you know the GA announcement's going to be coming somewhere in the six-month timeframe.

Yeah. That's totally accurate. Four months.

Yeah. It's the cycle. So if they're announcing it in the first half of the year, they're going to GA it in the second half.

If they announce it in the second half, it'll be GA first part of the next year. So that's the kind of, "Okay, we did what we said we're going to do. We gave you enough to work with to get us some feedback to be able to get to GA.

" That's sort of the next question. You were talking about MCP. Is that what it's going to be, or is it something more?

Is that the end state? Right. Exactly.

What does that look like? But I actually want to go to the drop because I'm going to spend a little bit more time on next segment. So we're going to go to the drop.

Okay, it's time for the drop. A lot of tools are coming out. There's meta-prompting tools.

Yes. There's CMax, to be able to have multiple sessions and control dashboards for all of your agentic operations, to manage and instruct as you become the engineer of agents driving software development. One of them, I think one that you came across, it had a really interesting...

Getting something done. I don't remember what the name of it is. That's close.

Getting stuff done. I'm not even sure we can really say that. Or your producer might- We're going to say it once.

Getting s**t done. Okay, now we said it. Wow, okay.

You went there. All right. Mitch.

It's out there. Okay, so we've just lost half our audience. We'll call it GSD from now on.

The rating just went from PG-13 or 14 to R, so we're okay. Whatever. Sorry, folks.

So that's another tool. Well, talk about what it is. Yeah.

Because it has kind of an interesting take on it, the whole prompting. It is. It's funny because the fellow who built this built it because he felt that some of the spec-driven development tools that we were using, such as Superpowers, which you and I have chatted about.

We need to do this, Matt. We need to sit down and actually go through it and start and use it a little bit, so the folks listening and watching who maybe haven't seen it can see what it does. But anyway, he said, "Yeah, that's a little too complicated.

" And- Yeah ... by simplifying, he meant to create a meta-prompting, context-engineering, and spec-driven development system. Which sounds- Sounds ambitious ...

a little complex to me, but it kind of makes me think one big thing, and that big thing is that is this not, and is Superpowers and all of the other spec-driven tooling we have, is that not a reflection of just how little we know about agentic development and how to use it effectively? I said this about OpenClaw when it first came out, I said this to Alan Shenkman. Well, he actually wrote an article.

I have to be careful what I say to him about the one line, one sentence I told him. " People will take that and create the next thing of it. People will come and create, get stuff done- Yep ...

using some ideas that came out of that. Maybe use some of it as a base. I'm not saying it is in this case.

But it can level up that way, or it can be just what spawns the next idea. And in a world where you can create so much more software yourself, and there's not a tool out there that does that, I'll do an open source project. Well, now I can do an open source project in a week- Right ...

and I can have something I am comfortable enough to put out to the world. That's a different dynamic. I mean, that is...

creature. It's not only bring your own tools, it's create your own tools. Yes.

It's an environment we're moving to. Yeah. It used to be BYOD or whatever you wanted to put as the last letter, and it's not bring, it's build.

Yeah. It really is. And it's not science fiction.

You can do that now. The stuff just in my kind of little world where I do some development to really help understand what this is all going to is, things happening in 30 minutes that would take, I don't know, three weeks before that- Well- ... to get through the whole process.

But that's because I'm slow But hang on, because some of these tools we're talking about, and I don't know if you want to call them just harness wrappers or CLI wrappers or whatever, actually slow you down quite a bit. Mm-hmm. Okay.

And is that a bad thing or a good thing? I don't know. Mm-hmm.

But if I, for example, for me, in my daily work, I use Conductor within the Gemini ecosystem at Google, and with that, you are sort of compelled to create a sort of spec of what you want to accomplish, and that is a discussion that happens. From that discussion, Conductor will generate a lot of material- Yeah ... which it will use to both stage and set the different steps that it will take to get to the outcome that you're talking about.

Also- Mm-hmm ... how to test for success and validity of what it builds on the way, which is vitally important, and one of the things I think these tools really do a good job of, and that is take off our plates, having to figure out how to test correctly . I don't know.

Sometimes maybe it's just a false sense of security. What do you think, Mitch? That we might hand some of that over to these tools.

But anyway, they could serve as almost a secret tax on your token budget because of that. Well, you're allowed to use those. Use local models.

Right. I think there's kind of two worlds we live in. " "Okay, great.

Here you go. Here's your snake game," or whatever. The single-shot prompt- Yep ...

nearly of creating by coding something. And then there's, "Well, I built all these things. " Yep.

"So for no other purposes, so when I restart another session, I can go back to where we are," right? Nothing else. It's kind of a memory mechanism.

But it's taking those things and doing the next incremental thing on top of it. Even if it's adjacent. It isn't just reusing stuff- No ...

it's actually part of what you take with you. It's institutional knowledge, is it not? Yeah.

Yeah. It is. " Right.

So now Mitch is doing it on his own, and he's running at a deficit there. " You did document your code, right? Okay.

No? He's self-documenting, obviously. Yes.

It's intuitively obvious. It must be you. It can't be my code if you can't understand it.

Yes, okay. Snark, snark. But anyway, I think to me, the point of the segment in using, whether it's getting stuff done or DataGoo or whatever it might be, is these tools are being reinvented, and they're going to be reinvented multiple times.

Yeah. That is the cool thing at the moment. Yeah.

Or that is the handy thing at the moment. But what helped us build the specs six months ago is going to look a lot different than it does now. Maybe it'll be the same- Because it's evolved ...

but maybe it'll be like just a re-envisioning of what we've come up with. Like maybe getting stuff done is overkill, but really good kill that should be applied in moderation or aspects of it, right? If I'm using open code, and I do very much appreciate just using plan mode and build mode.

And it's the same- Mm-hmm ... with CloudCode, the way they've set that up such that the framework, the harness kind of knows that when we're in plan mode, we're not trying to solve problems . We're trying to- Right ...

think about this thing. Hold off. Don't write code yet.

Right. Maybe that's enough, depending on, like you were just talking about, depending on how well documented the project is, how complex the project is, all that stuff. So it's not a one size fits all by any stretch, Matt.

I think the collapse of that is, or how it collapses is our and AI's ability to institutionalize or memorialize what we've done- Yeah ... why we've done it, and what the outcomes led us to, and what we learned from that. Mm-hmm.

Like in this work that I've done is everything is journaled. Whenever I learn something, that goes into the journal. Whenever the AI agents I'm using learn something, "Oh, that's an error we've seen before.

Here's what we have to do. " Okay, great. That goes into the journal.

So it's taking this corpus of knowledge, term I know you use a lot, and collapsing how we can use that. Yeah. Put that to work quicker in the process.

So when we go to do a specBut we're talking about building it all on these ways of doing things, but maybe rebuilt again- Right ... in a better way- Don't want to reinvent that- ... from what we did before ...

from start. Then- Yeah, it's pretty mind-boggling You know what you're talking about- It's interesting ... and that journaling is so interesting to me because, you know what that reminds me of is this concept that's probably lost to civilization now called a commonplace book- Oh ...

where you, like Balzac and Thoreau and other writers in the turn of the century, other century ago- Other century ... right, previous to the previous were writing. Centuries ago, centuries.

Exactly. Or would write down everything that they felt they might find useful later. " Tolkien.

Take Tolkien and all he wrote- Yes ... before he ever wrote anything. Exactly.

Right? There was such a vast amount of information he had created. He didn't use it all.

Matter of fact, he didn't use a lot of it, but he built on it. Right. It was built on by what he had written.

He picked what was relevant and important. Yeah. Yeah.

It led him to a place where he could tell that story. Yep. That's kind of what we're going through.

Feels like it. Very compressed. I'm with you.

Yeah. I need to get a commonplace book going. Insightful stuff.

There you go. " Well, it's time to be on the road again, I know you're doing some traveling, I think, to Google next. Yes, and you're off to- And yeah, I'll be- Probably as folks are watching this, you're at RSA Conference.

That should be interesting. RSA Conference, and then hopping on a plane to New York, hopping on a plane to Prague, hopping on a, I don't know, Las Vegas is always somewhere in the mix. Very true.

It is the Prague- Multiple times ... of America, yeah. It is, yes.

I think it's the hub of tech conferences, at least one of them. Mm-hmm. So, well, safe travels, my friend.

Same to you. It's been good doing another episode. And everybody, thank you for joining us.

We really appreciate you guys listening in and love your feedback. Tell us what you'd like us to look into or talk about. As Mitch said earlier, we get pummeled with information daily and we'd love to dig into whatever is of interest to you.

That would be great. We'd love to hear from you. We'd love to hear about what you're doing, too.

Tell us. We can learn from you. Yep.

So, all right, everybody. Safe, whatever you're doing, travels, coding, data engineering, whatever it might be, and continue to learn. Stay thirsty.

Keep learning. Stay hungry. All right, everybody, take care.

We'll see you next time. Control, this is Agent Dev. I'm in position.

Copy that, Dev. Standby for go. Standing by.