GS

Gábor Szárnyas

Things Gábor Says on Podcasts

Where to Find Them

Gábor Szárnyas has been a guest on The New Stack Podcast .

Recently: “DuckDB just shipped its biggest release ever. Its own team says not to run it in production just yet.” on The New Stack Podcast (November 2026).

What They Said

“If you asked me what I think about them maybe a year ago or even six months ago, I would have said that they were great, but you need to have your data in an openly available format … where the other agents can read it. But it seems that this is no longer really a problem. The agents are quite apt at reading proprietary databases and reading proprietary file formats, maybe at the expense of efficiency. But it seems to me that these silos are now breaking down.” — Gábor Szárnyas, The New Stack Podcast

Asked how his view of coding agents and data has changed in a year, he says agents now cope with closed formats that used to push teams toward open ones. It is a claim that data lock-in is weakening.

The New Stack Podcast · 2026-11-16 Permalink → Listen →
The New Stack Podcast Around 31:55 into the episode
Gábor Szárnyas

…into your face and says this is not right, and it's usually not very helpful. And we have been working on improving error messages, and this is already in a Git branch of DuckDB, and this will ship with DuckDB 2.1, and this will make that iteration cycle that we already talked about even shorter.

Jennifer Riggins

…have always been a problem. So, I think that's a really good idea. And boring use cases is where we want to fix things and where we scale our AI and our human developers for sure. So, on the personal front, how has your view of making data tools usable for coding agents changed over the last year?

Gábor Szárnyas

So, I think coding agents these days are really impressive. They keep impressing me with what they can do. And I think that they are starting to break down these silos, these data silos to some extent. So, if you asked me what I think about them maybe a year ago or even six months ago, I would have said that they were great, but you need to have your data in an openly available format, maybe in an open lake house format or in an open database where the other agents can read it. But it seems that this is no longer really a problem. The agents are quite apt at reading proprietary databases and reading proprietary file formats, maybe at the expense of efficiency. But it seems to me that these silos are now breaking down.

Speaker 4

Which is the best way to then come to fruition on your company data, where you actually, instead of open like a ChatGPT or a Gemini that's not trained on your internal systems, will not be as interesting. Your unique value proposition is on your internal systems and your internal data. So, to safely

Jennifer Riggins

bridge those silos, that's a fantastic use case. So, now putting on your prediction fortune teller hat by next November 2027, what do you think will be true about how developers and coding agents work with data? That isn't necessarily true today.

Gábor Szárnyas

I think with all these breaking down silos, it's already happening among many industries, but it's not really there yet for regulated industries, finances, insurance, healthcare, and so on. And that's because you need to be really precise there. You need to be explainable for what you do and for…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“The agents I tend to think of as like a bat flying in with bad vision and then doing echolocation, just doing maybe hundreds more queries and then trying to make sense of the data. And every now and then, some of these queries tend to be quite expensive. So if you run it in a cloud database, your bill may be a fair bit higher at the end of the month than you would expect.” — Gábor Szárnyas, The New Stack Podcast

The host asked what makes DuckDB useful for AI agents. Szárnyas compares how agents explore data with how people do. He ties it to cost, since running locally keeps the bill predictable.

The New Stack Podcast · 2026-11-16 Permalink → Listen →
The New Stack Podcast Around 23:57 into the episode
Gábor Szárnyas

…slash library, we do list a lot of projects that are done in an academic or research setup. And it's papers that ship a DuckDB implementation. And we found that most often than not, academics are really happy to implement their ideas in DuckDB and then ship it as a DuckDB community extension.

Jennifer Riggins

…new CEO, Andy Jassy, who was talking. And he talked about DuckDB and how it could help AI agents query data. So what makes DuckDB unique in this time of great change that is useful for agents? And then if it's useful for agents, typically it's useful for developers too. So what is that benefit?

Gábor Szárnyas

…of a few things. One is this simple install and easy setup that's beneficial for agents because they can run it very quickly. The other is this predictable cost approach. If you query something that is a database operated by someone else who charges some price for you for your queries, that can be a problem because the way these agents work is ultimately a bit different from how humans work. The agents I tend to think of as like a bat flying in with bad vision and then doing echolocation, just doing maybe hundreds more queries and then trying to make sense of the data. And every now and then, some of these queries tend to be quite expensive. So if you run it in a cloud database, your bill may be a fair bit higher at the end of the month than you would expect. With DuckDB, this will run maybe on your laptop or your local instance. So the cost remains predictable. And finally, I think the third thing for this is that it used to be a problem that the agent prototyped something for you and then you needed to move it to another maybe proprietary database system that is in the cloud. And people try to work around this with all sorts of clever approaches…

Jennifer Riggins

Okay, fair enough. And now that you've come back, listeners, from Googling Bat versus Duck, which I will do later to watch those videos, it's very interesting. Let's talk about the DuckDB CLI command line interface agentic mode. How does that work with data for coding agents? How do you?

Gábor Szárnyas

I think the CLI is turning this on, as we previously discussed, by default, whenever it detects that it's running within a harness of a coding agent. And it also, with this front-loaded information package that you see, helps quicker iteration. And ultimately, we have to acknowledge that software…

Jennifer Riggins

Okay. So as we referenced when I called Andy Jassy, your new CEO, he is. But Duck Labs joined Amazon last quarter. So I think those that are users and invest in the open source community, including your academic partners and friends, they kind of want to know what does that mean for the open source…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

Collections They Appear In