By Data pipelines built for human analysts are misaligned with the systems now consuming them. Agents retrieve continuously, operate autonomously, and require data shaped to their access patterns, not dashboards. Most existing ETL infrastructure was never designed with that consumer in mind.
In this interview on TFiR, Michel Tricot, Founder and CEO at Airbyte, breaks down what genuine architectural rethinking for agentic consumption looks like, why it must begin inside the organization before it reaches the product, and how Airbyte‘s internal Hydra project is reshaping connectors at the source.
Guest: Michel Tricot, Founder and CEO at Airbyte
Show: TFiR
Here is what every data engineer and platform architect building for agentic workloads needs to know.
Technical Deep Dive
Q: How can you tell whether a vendor has genuinely rebuilt for AI or just relabeled an existing product?
Michel Tricot, Founder and CEO at Airbyte, argues the signal is internal, not external. A vendor that has genuinely prepared for AI will have transformed how it builds its own systems before changing what it ships. If the only evidence of AI-readiness is a chatbot or a badge on the marketing site, the underlying architecture has not changed.
“As an organization, it has to start with you, not with the label that you put on the product. The moment you understand how powerful and how to tame that massive tool you have in front of you, the better you will be able to understand what needs to change within your product.” — Michel Tricot, Founder and CEO, Airbyte
Q: What specifically did Airbyte change about its architecture to support AI workloads?
Tricot identifies the primary consumer of data infrastructure as the determining factor. Because Airbyte’s connectors now serve agents as the primary consumer, the shape of the data they produce, how it is structured, indexed, and surfaced, must match agentic access patterns rather than human analyst workflows. This required rethinking connector design at the foundation, not adding a layer on top.
“The consumer of the product is an agent primarily. That is why we can say it is an AI product, because it is built for an agentic UX and an agentic experience.” — Michel Tricot, Founder and CEO, Airbyte
Q: What is Airbyte’s internal Hydra project and what problem does it solve?
Hydra is a fully agentic internal system at Airbyte that ingests open source contributions, Zendesk support data, and Sentry error data to determine how connectors need to be redesigned. It operates autonomously to reshape connector behavior so that outputs are appropriately structured for both agent consumption and traditional analytics consumption. Tricot describes it as a material internal investment, not a small project.
“It is a fully agentic system that takes all the open source contribution, all the Zendesk information, all the Sentry information and is actually reshaping how connectors need to work so that it is appropriately designed for agent consumption, but also for analytics consumption.” — Michel Tricot, Founder and CEO, Airbyte
Q: Does AI eliminate the need to physically move data in a pipeline?
No. Tricot is direct on this point: data must still physically move from source to destination. The LLM or agent consuming that data still requires it to land somewhere accessible. AI changes who and what consumes the data, and therefore how it should be shaped, but it does not remove the fundamental requirement for reliable data movement.
“At the end of the day there is a physical thing that needs to happen. Bytes need to move from point A to point B. This is something that AI cannot really replace because the data at some point needs to land on the LLM.” — Michel Tricot, Founder and CEO, Airbyte
Q: Why is search the core primitive in agentic systems and what does that mean for data infrastructure?
Tricot frames agentic systems as fundamentally search-driven. Agents do not browse or navigate; they continuously issue search queries against available data and then make decisions based on results. This means data infrastructure must be designed to support high-frequency, reliable retrieval. If the pipeline does not produce data in a form that agents can search effectively, the agent fails regardless of how capable the model is.
“Search was so important in an agentic world. At the end of the day it is search, because an agent needs to be able to always search and that is the only thing they do. They search and then they make a decision.” — Michel Tricot, Founder and CEO, Airbyte
Resources and Documentation
- Airbyte, open source data integration platform for moving data from any source to any destination
- Airbyte on GitHub, source code, connector catalog, and community contributions
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: These days a lot of vendors, they just slap AI ready label on, you know, existing solutions EXE tool. How do we know that AirByte is not doing the same thing? How, what specifically did you do to rebuild or rethink Airbyte’s architecture to make it truly AI ready versus just a label?
Michel Tricot: So that’s a very good question. And you know, when we talk about people putting an AI label on top of a product, the first thing is as an organization, it has to start with you, not with the label that you put on the product. It means like how do you enable yourself first with AI? Because the moment you understand how powerful and how to tame that massive tool that you have in front of you, like the better you will be able to understand what needs to change within your product. So it’s not just about adding an AI feature that is just adding a chatbot and answering questions for you. It’s really about how you are designing your own internal systems to power your product. And when we’re thinking about Airbyte, of course the end of the day there is a physical things that needs to happen. Bytes needs to move from point A to point B. This is something that AI cannot really replace because the data at some point needs to land on the, on the LLM. The, the piece here is the consumer of the product is an agent primarily. And that is why we can say like it’s an AI product is because it is built for an agentic UX and an agentic experience. The way we build the different connectors, we like to joke that we have a little, not little, it’s actually pretty big a project internally that we call Hydra, which is a fully agentic system that takes all the open source contribution, all the Zendesk information, all the sentry information and is actually reshaping how connectors need to work so that it is appropriately designed for agent consumption, but also for analytics consumption. But for me it’s really about thinking who is going to be the main consumer as an infrastructure product to be clear of the data that we’re providing and making sure that it is shaped in that way and that, you know, who knows, we knew that search was so important in an agentic world. Like it’s search, but at the end of the day it searched because an agent needs to be able to always search and that’s the only thing they do, they search and then they make decision.
