By Deterministic data pipelines gave engineering teams predictable, auditable control over what gets written to production systems. Agentic AI removes that predictability and puts write access into non-deterministic reasoning loops that can be triggered by any employee at any time. The governance frameworks built for small data teams do not scale to org-wide agent access, and the traceability requirements for agent-initiated writes are fundamentally different from anything reverse ETL workflows previously demanded.
In this interview on TFiR, Mario Moscatiello, VP Growth at Airbyte, breaks down how bidirectional data access changes enterprise agent architecture, why data lineage becomes non-negotiable in agentic pipelines, and how Airbyte is positioning its MCP integrations across frontier model providers.
Guest: Mario Moscatiello, VP Growth at Airbyte
Show: TFiR
Here is what every data engineer and platform architect needs to know.
Technical Deep Dive
Q: What is reverse ETL and how does it relate to agentic data write-back?
Mario Moscatiello, VP Growth at Airbyte, explains that reverse ETL, also called data activation, is the practice of transforming data in a warehouse and writing it back into operational systems like Salesforce. A common example is running a daily dbt model that identifies top spenders and automatically updates CRM records each morning so sales teams can act on that data. This pattern has existed for years before AI agents entered the picture, meaning the write-back infrastructure is not new but the actor initiating the write is now fundamentally different.
“You could have a dbt model that runs every day and says, who were my top spenders yesterday? And what you want to do is have a system that every morning writes that back into Salesforce so that the sales team can look at who were the biggest spenders and try to upsell them.” — Mario Moscatiello, VP Growth, Airbyte
Q: How does agentic write-back differ from traditional deterministic data pipelines?
Traditional reverse ETL pipelines are written by data engineers, follow deterministic logic, and run on a schedule. When an AI agent writes back to a system, the reasoning behind that action is non-deterministic, meaning the same inputs may not always produce the same output. Moscatiello emphasizes that this shift requires organizations to move from trusting code to trusting agent reasoning, which introduces hallucination risk where the agent could write incorrect or unauthorized data to a production system.
“When you go from a data team writing very deterministic code that is going to run every day and just update stuff in a CRM, to an agent writing back to a system, you need to make sure that the agent is not hallucinating and is writing the right data to the right system.” — Mario Moscatiello, VP Growth, Airbyte
Q: Why does org-wide agent access create a governance scale problem that data teams are not prepared for?
In a traditional data-mature organization, the number of people who can initiate writes to production systems through data pipelines is small, typically 50 to 100 people on a dedicated data team even at a company of 10,000 employees. When every employee’s AI agent can trigger write-back operations, the access surface expands by orders of magnitude. Moscatiello notes that this scale shift means organizations can no longer rely on team-level access controls and must implement per-user, per-agent authorization checks to determine whether a specific person is even permitted to write to a specific record.
“You go from a few people in an organization owning and governing access to data to suddenly everybody being able to work on data. The scale at which that happens is just massive, and so it introduces a lot of challenges.” — Mario Moscatiello, VP Growth, Airbyte
Q: What data lineage and traceability requirements are specific to agentic pipelines?
Because agent reasoning is non-deterministic, organizations need audit infrastructure that can answer specific questions after the fact: why did the agent take an action, what data did it use to reach that decision, who or what initiated the agent run, and whether the trigger was a schedule or a human prompt. Moscatiello frames this as a traceability requirement that is more demanding than anything traditional lineage tools were designed for, because the action itself cannot be fully predicted from the inputs alone.
“You need to be able to understand why did an agent take an action. You need traceability. This is a non-deterministic flow where an agent is reasoning and taking action, but you always need to be able to trace that back into why did the agent actually make the decision.” — Mario Moscatiello, VP Growth, Airbyte
Q: Where is Airbyte’s MCP integration available and what is the roadmap for other frontier model providers?
Airbyte’s MCP integration is currently available on the OpenAI Marketplace. Moscatiello confirms the company is actively working with all frontier model providers, including Anthropic, to get resources officially listed on their respective marketplaces. He notes that marketplace listing is a process that varies by provider, and that users can already install the Airbyte MCP into Claude directly by pasting a link without waiting for official marketplace placement.
“We’re working with all of the frontier labs to have these resources officially on their marketplaces. You can still go ahead and install an MCP into Claude. It’s just pasting a link.” — Mario Moscatiello, VP Growth, Airbyte
Resources & Documentation
- Airbyte, open-source data movement platform supporting reverse ETL, data activation, and MCP-based agentic integrations
- Airbyte Documentation, official documentation covering connectors, pipeline configuration, and platform features
- Airbyte GitHub, open-source repository for connectors and core platform code
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: Can you talk about, you know how it just kind of pushes your byte beyond just retrieving data into actually now you are not just looking at it as not just one way traffic, you are actually also taking actions like writing back to very sophisticated system like Salesforce. How does that bidirectional access change the way enterprise agents are architected and and what new governance challenges come with it? Because if you do allow your agents to start taking actions to start writing to your surfacing system, you need to have right governance as well. So it is not as that simple. So can you talk about the working of the cli?
Mario Moscatiello: I think the CLI is one component but where we allow developers to just interact with our platform directly from the terminal. But I think what we’re talking about here is more like us enabling agent to write back to systems. If we look back at the last five years, this has already happened when it comes to just pure data. So I’ll back up a bit. But essentially we also have this feature, but there are a lot of companies that a bunch of companies that have developed this feature before, which is called what we call reverse CTL or data activation. And what that means is that this is before AI. You could have a DBT model that runs every day and says like hey, who were my top spenders yesterday? And of course what you want to do is have a system that every morning writes that back into Salesforce so that the sales team can look at who were the biggest spenders and try to upsell them or trying to lock them into a contract as an example. And so this reverse etl, actually transforming data and writing it back to systems is not something that is just out today. It’s something that companies have been doing for a while. I think the issue is that when you go from a data team writing very deterministic code that is going to run every day and just update stuff in a CRM or in whatever system. When you go from that to an agent writing back to a system, first of all you need to make sure that the agent is not hallucinating and then the agent is writing the right data to the right system. And second, you go from a few people within organization, even if it’s a large organization you can team off. Even A company of 10,000 people is going to have maybe 50 people, 100 people in the data team if it’s already very data mature. When you go to from like 50 people to suddenly everybody’s agent having access to writing back to systems, how do you know that a specific person is even allowed to write back to the record? Or how do you know that a specific person has access to the information he needs for the agent to take action? And so from a governance standpoint, it requires way more robust governance and so on and so forth. But you also need way better data lineage because you need to be able to understand, like why did an agent take an action? You need traceability, you need to understand, yes, this is a non deterministic flow where an agent is reasoning and taking action, but you always need to be able to trace that back into, okay, but why did the agent actually make the decision? Why did the agent wrote that data? Who initiated that agent? Was it on a schedule? Was it automatic? And so on and so forth. So I think you’re going from a few people in an organization owning and governing access to data to suddenly everybody being able to work on data. And so the scale at which that happens is just massive. And so it introduces a lot of challenges.
Swapnil Bhartiya: So it’s available on OpenAI Marketplace. What about other like cloud and Tropic cloud is there?
Mario Moscatiello: Yeah, we’re working with.
Swapnil Bhartiya: What is a long term?
Mario Moscatiello: Yeah, we’re working with them. You know, it’s, it’s a process. Like every company. OpenAI was quick in this case, but it’s a process. And we’re working with all of the frontier labs to have these resources officially on their marketplaces. You can still go ahead and install an MCP into Claude. It’s just pasting a link. Of course, it’s nice to be in the marketplace so that people can find it quicker and their team is also doing some work on it. But we do work with all the model providers.
