By Enterprises are deploying AI agents through multiple interfaces simultaneously. Marketing teams query through MCP, developers build with SDKs, and engineers operate from the CLI. When each interface pulls from a different data source, the same business question returns different answers depending on who is asking and what tool they are using. That inconsistency destroys trust in AI outputs and creates a governance crisis that compounds as agent deployments scale.
In this interview on TFiR, Mario Moscatiello, VP Growth at Airbyte, walks through how shared data infrastructure resolves agent output inconsistency, why model-agnostic architecture is becoming a survival requirement for enterprises, and what token budget overruns reveal about the state of AI governance in large organizations.
Guest: Mario Moscatiello, VP Growth at Airbyte
Show: TFiR
Here is what every platform engineer and enterprise AI architect needs to know.
Technical Deep Dive
Q: Will one AI agent interface standard win out, or will enterprises have to manage MCP, SDK, and CLI simultaneously?
Mario Moscatiello, VP Growth at Airbyte, says no single interface will win because different roles inside an organization require different tools. A marketing team using Claude or OpenAI will naturally gravitate toward MCP as their working interface, while developers building custom agents will reach for the SDK, and engineers operating from the terminal will use CLI. Moscatiello argues the interface layer will remain fragmented by design, and the right engineering response is not to force consolidation at the interface level but to enforce consistency at the data layer beneath all of them.
“Different people within an organization will use different interfaces for different types of work.” — Mario Moscatiello, VP Growth, Airbyte
Q: What happens to data consistency when multiple agent interfaces query enterprise systems independently?
Moscatiello frames this as the core enterprise AI data problem: if a marketing agent answers a revenue question through MCP and an engineering agent answers the same question through an SDK, the answer must be identical because it is the same business fact. If the underlying data infrastructure is not shared, the answers diverge and the business loses confidence in both agents. Airbyte’s position is that the interface a team uses is their choice, but the data those interfaces read from and write to must be a single, consistent layer across the organization.
“Whether it’s a question asked from an MCP or from an SDK, the answer should be the same.” — Mario Moscatiello, VP Growth, Airbyte
Q: How much control do developers actually have over AI tooling and model selection in enterprise environments?
Moscatiello draws a clear line between company size and developer autonomy. Smaller companies allow developers to experiment freely across providers with usage-based pricing. As companies scale, cost overruns force top-down consolidation. He cited an example of a large organization that exhausted its entire annual token budget within the first few months of the year, triggering executive-level governance conversations. At that scale, the board and C-suite negotiate enterprise contracts with model providers, and individual developer choice gives way to centrally mandated tooling.
“Larger companies will tend to have more top-down governance because these models are expensive and they need to track ROI.” — Mario Moscatiello, VP Growth, Airbyte
Q: Why is model-agnostic architecture becoming a competitive requirement for enterprises?
Moscatiello explains that the model landscape is moving fast enough that any enterprise locked into a single provider risks losing competitive advantage when a better model ships. If an organization builds its entire agent stack on one provider’s model and a meaningfully better alternative becomes available in three months, the ability to swap that model in should be an architectural given, not a major engineering project. He frames model-agnostic design as the balance point between cost control, which pushes toward consolidation, and flexibility, which demands optionality.
“If another provider comes up with a better model that is way better for you, you should be able to swap everything and just use that model because it will give you a competitive advantage.” — Mario Moscatiello, VP Growth, Airbyte
Q: What role are open source models playing in enterprise AI stacks, and how does model sovereignty factor into procurement decisions?
Moscatiello sees open source models gaining serious traction as enterprises look for ways to reduce vendor dependency and maintain data and model sovereignty. He noted that strong open source models are already emerging and that enterprises are beginning to evaluate model sourcing through the same sovereignty lens they apply to data residency. He also flagged the competitive pressure from Chinese AI labs, which he described as producing impressive models, and expressed a view that US leadership in open source model development matters for the broader enterprise market over the next five to ten years.
“Companies are going to look more and more at sovereignty, even with models. It is going to be an interesting five to ten years ahead.” — Mario Moscatiello, VP Growth, Airbyte
Resources & Documentation
- Airbyte, open source data integration platform for building agent-ready data pipelines across MCP, SDK, and CLI interfaces
- Airbyte Documentation, official docs covering connectors, deployment, and data infrastructure configuration
- Airbyte GitHub, open source repository for connectors and platform contributions
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: Of course, if you look at AirByte and I have been covering AirByte from the very early on, so I have seen your journey. Now, you folks are giving agents multiple ways to interact with Enterprise Systems, NCP, SDK, now CLI and of course availability in OpenAI App Marketplace. And it will be available in other places as well. How do you see these interfaces evolving? The AI market is changing very fast, so it’s hard to say that, but just, you know, based on your interaction, based on your own usage. Because today everybody is using AI either way, will you do you see one will win out as a standard or there will be a fragmentation, there will be a mess where enterprises will have to deal with all of them depending on the kind of agents they need. Also, a lot of players, even anthropic, they have launched managed agents because. Because they all understand that agents are becoming another big part of the problem for enterprises to manage.
Mario Moscatiello: I think you hit the nail in the head, right? I think it’s different people within an organization will use different interfaces for different types of work. I think if you’re in a marketing team and you’re using Claude or OpenAI and you want to use an MCP because that’s the interface, that’s your interface for work, that’s great. If you’re a developer and you’re building an agent and you need the data SDK and you’re using the SDK, that’s great. You’re building a custom agent for a specific thing, or you’re building an application that has agents in the back, that’s great. And if you’re again an engineer working out of your terminal and you want to have access to a command line interface to manage everything you do in this agentic application, that’s okay too. And I think that what’s the interesting thing is that something like we do with everyday agents is we can say, look, we know that enterprises need access to different interfaces for different classes of agents, but what shouldn’t change is the data infrastructure behind them. So that if your marketing team is using the MCP and they’re querying data, or your developers are using the SDK and the cli, they should read from the same data. Like if you know, at the end of the day the marketing team is marketing person is building an agent, that should answer how much revenue did we do yesterday? And an engineer doing it with the mcp and an engineer is trying to do the same with the SDK because they’re building a revenue agent to that question, like the agent Whether it’s, you know, a question is asked from an NCP or from an SDK, like the answer should be the same. And so what we can say is, look, it doesn’t matter whether your teams are querying from RMCP or building something custom with the SDK or the cli. We want to make sure that they’re reading and writing from the same data so that everybody in the business speaks the same language.
Swapnil Bhartiya: How much say developers have when it comes to AI versus how much is the top down? Because that also dictates the kind of framework you are building, the kind of tools you’re using, the kind of AI you’re allowed to use or not use.
Mario Moscatiello: I think in smaller companies developers have the freedom to experiment. Usually, you know, smaller companies are more flexible and they tend to have, you know, more usage based pricing when it comes to like, yeah, you can, you can experiment with a bunch of different providers and so on and so forth. I think when, when the companies like start scaling, you have, you know, companies like Uber saying wait on a second, like it’s, I don’t remember the exact month but they were saying hey, it’s March or April and we’ve already used all of our token budget for a year. We need to put some governance in. And so in that sense if you have, you know, the board and the executives, they probably go and negotiate with one of the providers and say hey, like if we were to deploy your models across the entire companies, what is the pricing? And so I think from a cost perspective, larger companies will tend to have more top down, you know, in that sense because these models are expensive and these providers like are expensive and so they need to put in, in place. Again we go back to governance. Like it’s not only governance for the data side but it’s also governance for how much they’re spending and how many tokens they’re using and are they getting roi? But I think what’s happening right now is that a lot of these companies are realizing that they also need to stay open to be working with different providers and building all of their stack in what we call like a model agnostic way. Because you know, for a company, if in three months, if you’re working with one provider and another provider comes up with a better model in three months that is way better for you. You should be able to in theory swap everything you’re doing and just use that model because it’s going to give you a competitive advantage. And so like, I think it’s kind of like a fine balance in between cost in between cost and sort of like flexibility. And that’s what we’re also seeing a lot of the open source models really, really taking shape. And you have some great open source models that are being developed. I certainly hope that the US Will start leading the charge with open source models because I think that we’re lagging behind, especially compared to China. They have some amazing labs, they’re coming up with amazing models. And I think that companies are going to look at more and more, are going to look at sovereignty, even with models. So it’s going to be an interesting, like five, ten years ahead.
