By AI agents operating on outdated data make outdated decisions. As enterprises push beyond chatbots into autonomous decisioning and chained multi-agent workflows, data freshness, discovery speed, and access latency have become first-class infrastructure requirements. Most current architectures were not designed to meet them.
In this interview on TFiR, Michel Tricot, Founder and CEO at Airbyte, covers how enterprises should define latency targets for agentic systems, how a context store addresses discovery speed and data freshness simultaneously, and how organizations at different maturity levels can structure data access for production AI agents.
Guest: Michel Tricot, Founder and CEO at Airbyte
Show: TFiR
Here is what every platform engineer and AI infrastructure team needs to know.
Technical Deep Dive
Q: What latency and data freshness requirements are enterprises setting for AI agents that need to act in real time?
Michel Tricot, Founder and CEO at Airbyte, says latency for AI agent decisions is fundamentally a business decision, not a fixed technical threshold. The realistic target today is to match the decision-making speed of a human first, then iterate down from there, because current systems still require LLM inference time and lack the full control infrastructure needed for sub-second autonomous action. On data freshness, Tricot is direct: if an agent is running, it must be looking at the most recent piece of data available, without exception.
“The thing that is for sure happening is that you can make this decision faster. Now you don’t have to go to sub 10 second or sub millisecond type of decision speed because the system today, it’s slow.” — Michel Tricot, Founder and CEO, Airbyte
Q: How does a context store balance discovery speed and data freshness for agentic workloads?
Tricot explains that the context store is designed to provide very strong latency on the discovery side, meaning an agent can find the relevant data quickly. However, the context store alone does not guarantee the agent is acting on the latest record. To solve this, Airbyte’s architecture gives the agent the ability to reach back to the source system directly and verify that no difference exists between the cached context and the live data before making a decision.
“If you want to make sure that you always have the latest version of the piece of data that the agent has at hand, you always have the ability to just look at the source system and make sure that there is no difference between these two pieces of data.” — Michel Tricot, Founder and CEO, Airbyte
Q: Are enterprises using agentic data platforms for autonomous decisioning and workflow orchestration, or just chatbots?
Tricot uses a self-driving car maturity model to describe how companies adopt agentic data infrastructure. Companies move through levels, starting with MCP-connected chatbots integrated with models from Anthropic, OpenAI, or Perplexity, and progressing toward fully automated agent processes built using open source SDKs or APIs that expose data access primitives directly. At higher maturity levels, teams build complex chained workflows that combine predictable and less predictable tasks, with gates that enforce correct execution direction across multiple agents.
“As you go and automate more, you also have the ability to just build more complex workflows that are chaining different agents, adding predictable and a little bit unpredictable tasks with the right gate to make sure that the workflow is executing appropriately.” — Michel Tricot, Founder and CEO, Airbyte
Q: What data access primitives does Airbyte expose for teams building automated agent pipelines?
Tricot describes the agentic data platform as an infrastructure product that solves three core problems: data access, data discovery, and data write. For teams building specific automated agents rather than chat interfaces, Airbyte provides an open source SDK and APIs that expose these data access primitives directly. This allows engineering teams to construct agent logic that queries the right information at each step of an automated process without relying on a chat layer as the intermediary.
“If you are looking to build a specific agent that is going to run as an automated process, this is where you can use either the open source SDK or APIs that give you access to these data access primitives.” — Michel Tricot, Founder and CEO, Airbyte
Q: How should infrastructure product teams account for different levels of enterprise AI maturity?
Tricot argues that any company building an infrastructure product for agentic AI must design for the full spectrum of customer maturity, not just the most advanced use cases. Airbyte’s approach is to hand-hold companies through the adoption journey, ensuring teams can start where they are and progress without being forced into architectural decisions they are not ready for. The platform’s modality supports entry-level MCP integrations through to fully orchestrated multi-agent systems, so customers are not blocked by their current maturity level.
“Every company that is building an infrastructure product needs to take into account the level of maturity of companies and make sure that they are handholding all these companies to make sure that they have the most successful experience introducing AI into their system.” — Michel Tricot, Founder and CEO, Airbyte
Resources & Documentation
- Airbyte, agentic data platform providing data access, discovery, and write primitives for AI agents and automated workflows
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: Since we are talking about, you know, dynamic data, not steady data, what kind of latency or freshness requirements are enterprises demanding or asking for when they are building agents that need to act in real time and you are the one, you know, providing them with the data.
Michel Tricot: In terms of latency, I think it’s a business, it’s a business decision like how fast you want specific data decisions to be made. The thing that is for sure happening is that you can make this decision faster. Now you don’t have to go to sub 10 second or sub millisecond type of decision speed because the system today one, it’s slow. I mean you need to, you need to let the LLM run and come up with a solution. And also we don’t have yet in the overall world all the control systems in place. And so I think what’s going to happen is first let’s get to the latency of decision making that a human has and just decrease as you were learning how to make the system work and just iterate and get that latency down. Now in terms of data freshness, well, if you have an agent running, you want it to be looking at the most recent piece of data. So that’s why when we’re thinking about the context store, the context store to give you latency, very strong latency when it comes to the discovery. But if you want to make sure that you always have the latest version of the piece of data that you, the agent has at hand, you always have the ability to just look at the source system and make sure that there is no difference between these two pieces of data and make decision based on the most recent data record that you have? Does that answer your question?
Swapnil Bhartiya: Can you also talk about how are customers using this? Are you seeing that customers are using it for agentic workflows that go beyond chatbots like autonomous decisioning, workflow orchestration or closed loop system? What does this look like in practice?
Michel Tricot: If you have seen when we’re thinking about the agency platform, we’re always thinking about the modality by which people access data, by which agents access data and also the maturity of adoption. And you know, we always, you know the same way we’re talking like for self driving cars in like level 1, level 2, level 3, level 4, level 5. Most companies follow a very similar, a very similar motion across these levels. And the agentic data platform is an infrastructure product meaning that it solves the access, it solves the discovery, it solves the write. Now how you consume it is going to be different based on your level of maturity or the problem at hand and having that as part of your chat. Well, this is something you can solve through an MCP and connecting it to anthropic OpenAI perplexity, whatever, whatever. Whichever bot that you want, what you want is just to make sure that as the LLM is progressing, it’s able to just query the right information. Now if you are looking to build a specific agent that is going to run as an automated process, this is where you can use either the open source SDK or APIs that give you access to these data access primitives. And as you go and automate more like, you also have the ability to just build more complex workflows that are chaining different agents, adding predictable and a little bit unpredictable task with the right gate to make sure that the workflow is executing appropriately and that you are guiding the different agents running on it into the right direction. So that is really how we’re thinking about data access. And every company that is building an infrastructure product needs to take into account what the level of maturity of companies and make sure that they are. We are handholding all these companies to make sure that they have the most successful experience introducing AI into their system, into their team, based on where they’re at in their journey.
