By Across the enterprise, AI agents are breaking in production. Development teams debug for weeks, swap in newer models, and rewrite prompts — yet agents still return stale answers, hallucinate facts, or fail to complete basic workflows. The instinct is to blame the model. According to Michel Tricot, Co-Founder and CEO at Airbyte, that instinct is almost always wrong.
The real failure point is upstream: a fragmented data layer that forces agents to chain five or six live API calls just to answer a single question. Each call burns tokens, adds latency, and risks pulling back contradictory results from systems that have never been reconciled with one another. By the time the agent formulates a response, its context window is already polluted. The model is doing exactly what it was designed to do — the infrastructure around it is not.
This is the production reality that Airbyte is now targeting directly. Known for building one of the most widely adopted open-source data integration platforms — with over 600 connectors and a catalog that spans Salesforce, Zendesk, Gong, HubSpot, Slack, and hundreds more — Airbyte has now extended its platform into the agentic era with the launch of Airbyte Agents and its core capability: the Context Store.
The Context Store is a pre-materialized, unified, and permission-aware context layer purpose-built for production AI agents. Rather than letting agents autonomously discover and fetch data at runtime — a process Tricot describes as slow, expensive, and error-prone — the Context Store pre-indexes structured and unstructured data from across an organization’s systems, resolves entity relationships between them, and surfaces exactly the context an agent needs, at the speed agents operate.
The launch marks a meaningful identity shift for Airbyte: from data movement platform to agentic data platform. It is a shift, Tricot argues, that the market itself has been demanding — quietly, through changed behavior in how both open-source users and enterprise customers have been deploying Airbyte’s connectors since early 2024.
The Guest: Michel Tricot, Co-Founder & CEO at Airbyte
Key Takeaways
- Enterprise AI agents fail in production primarily because of fragmented, unreconciled data — not model limitations. Agents pull too much irrelevant data, burn tokens, and corrupt their own context windows before returning a response.
- The Airbyte Context Store is a pre-indexed, permission-layered knowledge graph that gives agents fast, cross-system access to the right data — resolving entity relationships (e.g., “Walmart” in Salesforce = “Walmart” in Gong) without requiring agents to do the heavy lifting.
- Airbyte launched with 50 connectors prioritized around support (Zendesk), sales and revenue (Salesforce, Gong, HubSpot, Marketo), with 600+ connectors in the broader catalog.
- Write-back governance is enforced through explicit permissioning, action tracking, and eval loops — agents are given only the actions they are authorized to take, and every action is logged.
- Agent operations is Airbyte’s new unit of pricing — charged only on successful retrieval, read, or write actions — aligning cost directly with agent value delivered.
***
In this exclusive interview with Swapnil Bhartiya at TFiR, Michel Tricot, Co-Founder and CEO at Airbyte, discusses why enterprise AI agents fail in production, how the Airbyte Context Store solves the cross-system data retrieval problem, how identity resolution works across tools like Salesforce, Gong, Zendesk, HubSpot, and Marketo, the difference between MCP and SDK adoption patterns, write-back governance design, the new agent operations pricing model, and how Airbyte has evolved from an open-source data movement platform to a full agentic data platform.
Why Enterprise AI Agents Fail in Production
The most common diagnosis when an AI agent underperforms is that the model is at fault. Tricot challenges this assumption directly, drawing a parallel between the data infrastructure challenges agents face today and the challenges enterprises faced when standing up data warehouses for human decision-making. Both require a unified, 360-degree view of the organization — agents just operate at a speed and autonomy level that humans do not.
Q: Where have you seen enterprises fail when they build their data pipelines for agents?
Michel Tricot: “I think it’s not too different from what happened with warehouses. In order for a human to make a decision on top of data, you need to make sure that the human has access to a 360-degree view of what is happening in the organization, in their business. For agents, it’s not that different. It’s just that the way you consume data, the type of operation you’re making, is a little bit different and requires some new type of technology, especially when it comes to search. That is really what has been motivating — enhancing the capabilities of Airbyte to make sure that we provide access to data to these new types of consumer that operate at a speed that a human cannot, and that is required to make decisions almost autonomously.”
Q: Where exactly does the agent pipeline break — what is the specific point of failure that Airbyte Agents is solving?
Michel Tricot: “Data is fully, completely fragmented across different systems. When you let an agent autonomously discover this data, the first problem you have is the same one you have with anything data-related. You have fragmented systems — you need to be able to join data across these systems. I need to know that Swapnil on System A is the same Swapnil on System B. For an agent to be able to do that without the proper infrastructure, it requires the agent to pull a lot more data to be able to make that join. So that is the first thing — the ability to join is very hard. The ability to search is also very complex for an agent. If I’m searching for information about Walmart as a company, there are a lot of systems that don’t have this ability to search, and an agent needs to be able to discover the context before it can take an action.”
Michel Tricot: “That’s what we call the agentic data loop — you always discover, you fetch the data, and then you act on that data based on the outcome your agent is trying to provide. This is really a place where we see a lot of effort failing. Not having the ability to centralize and have access to that context breaks in production because these agents spend a lot of time just pulling data, pulling more data, pulling more data — data that is actually irrelevant to performing the task. That means you’re going to be using more tokens, you’re going to have pollution of your context, and you’re also going to have a lot more latency because you need to pull this data into the agent directly. That is really where we see a lot of things failing in production. In general, when you do a very simple demo, you can dump the data into a text file, copy and paste it, or let the agent just read the file — but that’s good for a demo. That’s not how agents actually operate, and that’s not how you want agents to operate in production.”
Models, Prompts, and the Real Source of Confusion
Tricot acknowledges that models are not entirely without fault — prompt adherence degrades as context windows grow, and rapid model iteration by providers like Anthropic and OpenAI forces enterprises to continuously re-validate their agentic workflows. His answer is architectural rather than reactive: move away from single-prompt, fully autonomous agent designs toward predictable, workflow-driven agentic systems with isolated decision points.
Q: How are you approaching the fact that even with good data, the model layer itself — prompt adherence, context length, rapid model releases — creates its own instability?
Michel Tricot: “In today’s world, companies need to adapt. We are at the beginning of a big platform shift. These models need to be monitored. The data they have access to needs to be monitored, and we will need to adjust how agents are performing. If you just look at what agents were able to do in coding a year ago versus the breakthrough that happened in the past six or seven months, people had to change the way they work — they had to change workflows. So I think this is more a cultural thing within organizations. The moment they want to adopt AI technologies, they need to be adapting to the massive changes happening in the market.”
Michel Tricot: “When we started to build agents, everything was coming from the prompt — we would let the agents decide when to do A, when to do B, when to do C. That led a lot of systems to just fall on their face, because you’re suddenly giving the responsibility to the agent to decide what the next step is. What we’re seeing with our customers now is that people are moving toward a more workflow and predictable workflow type of approach — they try to isolate the specific places where they need an agent, whether they need to pull data from the context or from an API, and have predictable branches in their workflow. That allows them to bring some certainty into something that is unpredictable. People are moving away from the naive approach of ‘I have a playbook and I just shove everything into my agent.'”
What Is the Airbyte Context Store?
The Context Store is the central technical product in the Airbyte Agents launch. It is not a RAG pipeline replacement, nor is it simply a vector database abstraction. It is a pre-materialized, permission-enforced knowledge layer that indexes both structured and unstructured data across connected systems and surfaces entity relationships between them — enabling agents to retrieve exactly the context they need without over-fetching, without resolution errors, and without latency spikes from live API chaining.
Q: What is the Context Store, and what does the work of assembling context before the agent runs actually involve?
Michel Tricot: “The Context Store that we’re launching is really a pre-materialized, unified, and permission layer for agents to directly access live data. What is necessary when you want an agent to get your data is, first, it has to be indexed — whether it’s structured or unstructured data. We’ve talked a lot about RAG and vector databases. This is just indexing of data and offering the ability for agents to search. That is what the Context Store is actually doing. It is built on top of very reliable data movement primitives that Airbyte has built over the past few years, and we’re adding on top of it the ability to index that data and to make it searchable. But it’s also about what relationships exist between the different pieces of data — making sure that company Walmart is connected to this list of customers, that these customers are having this type of support ticket with Walmart, and so on. The relationship here actually matters a lot, because that’s how the agent is actually operating — it always starts with a first piece of information, and then it’s going to pull every single string it can to have the most relevant context. That’s really the primitive that the Context Store is bringing to these agents.”
MCP Endpoint vs. Native SDK: Who Reaches for Which
Airbyte has built two access patterns into the Context Store: a Model Context Protocol (MCP) endpoint for teams already building with MCP-compatible orchestration layers, and a native SDK for engineering teams building internal systems. Tricot’s framing of the distinction is notably non-segmented — both paths access the same underlying context layer, and the goal is uniform data quality across every agent in the organization, regardless of who built it.
Q: Who is reaching for the MCP endpoint versus the native SDK, and how are teams adopting each?
Michel Tricot: “The product is the Context Store, and it is really this global knowledge of everything that is happening within your company. Companies are going through a massive change right now — they want every single team to become AI-proficient. What we want to do is make sure that the technology we’re bringing to market is the same one, whether you are a RevOps person working on sales and marketing information, or whether you are an engineer building an internal system for your team. For us, it’s not so much about what the audience is — whether it’s a developer or a RevOps person. It’s more about providing a solution for companies to actually adopt Airbyte and make sure that every single person across the company, every single agent that is built, is operating out of the same context.”
50 Launch Connectors: How Airbyte Prioritized and How Identity Resolution Works
Airbyte launched Context Store with 50 connectors drawn from its catalog of 600+. The selection was deliberate: support functions (Zendesk) first — because support was the earliest enterprise LLM use case — followed immediately by revenue and sales functions (Salesforce, Gong, Marketo, HubSpot), where AI adoption is accelerating fastest. Entity resolution across these systems works by leveraging all available metadata on each indexed record and mapping it across systems — fuzzy-matching company names like “Walmart” across Salesforce and Gong regardless of minor formatting differences.
Q: How did you decide which 50 connectors to launch with, and what does entity resolution across systems like Salesforce, Gong, Zendesk, and Slack actually require under the hood?
Michel Tricot: “We went for the first 50 connectors based on identifying which problem we wanted to tackle first. When LLMs first emerged, the first place people invested was support — because support has a lot of context consumption, a lot of response creation. That was really a low-hanging fruit, and that’s why we’re supporting Zendesk and other support systems. The next one is really happening on the sales and revenue-generating functions. These are not always the most AI-forward teams, but the moment they’ve tested what AI can do for them, they are going to embrace it. That is the reason why Salesforce, Gong, Marketo, HubSpot are very, very important.”
Michel Tricot: “In terms of identity resolution, these teams have a subset of very well-identified entities that they support — customer, contact, opportunities, visitors. That is for us a way of providing fast identity resolution and linking these data sets together, because most of these systems operate out of the same entities. As we go to different sets of connectors, we will add more and more entities that are linked together. How we do it is we leverage all the pieces of information available in every single record that we index, and we provide the ability to just map these informations together. So if you’re on Salesforce and you have Walmart, and on Gong you also have Walmart — whether it’s written the same way, whether there are a few typos — that doesn’t matter for an LLM and for indexing. That’s how we make this type of link, plus all the additional metadata associated to it.”
Write-Back Governance: Permissioning, Action Tracking, and Evals
Giving agents the ability to write back into systems — updating CRM records, creating support tickets, adjusting campaigns — is where AI infrastructure moves from impressive demo to genuine operational risk. Tricot addresses this directly, arguing that the answer is not a perfect system that never makes mistakes (which does not exist) but a layered approach combining narrow permission scopes, comprehensive action logging, and eval loops that help agents learn from their errors over time.
Q: When you allow write-back functionality, how do you also build governance around it so agents are only doing what they’re allowed to do?
Michel Tricot: “They will hallucinate no matter what. I think we cannot expect to have a perfect system that doesn’t make mistakes — even humans, when they operate with these systems, make mistakes. What we have is really this ability, when you’re building agents, to enforce specific permissions. You’re not just giving access to Zendesk and every single endpoint available on Zendesk to write. Maybe the agents you have, because of the maturity of your AI adoption, should just write drafts. A draft has a very limited impact, and it’s fine if it makes a mistake — you can always have a human in the loop after that. What is important is building the ability to track every single action that has been performed. Not all of them you can revert, but with the right evals, you can make sure that the agent is learning on what was good, what was not good, and has that memory of good actions and bad actions.”
Michel Tricot: “In terms of a system of rules, this is not something we support today — this is something that needs to be encoded into the workflow when building agents. But for example, if you’re consuming the context directly via an MCP, the MCP has instructions to make sure that the operator of Claude or OpenAI is actually confirming that this is the right action to perform. For agents, it requires a little bit more work, but you can already get by with permissioning and selecting which actions you want your agents to be able to perform.”
Automations: A Visual Workflow Builder for Agentic Pipelines
Alongside the Context Store, Airbyte is launching Automations — a visual interface for building agentic workflows directly within Airbyte. Tricot positions it as both a fast on-ramp for new users and a philosophical statement about how agents should be designed: not as prompt-stuffed autonomous systems, but as structured, iterable workflows with predictable branching.
Q: What is Automations, where does it fit in your roadmap, and what is Airbyte today?
Michel Tricot: “Airbyte is an agentic data platform. It is really about connecting every single data system that you have and making it consumable. Automation is really a way for people to get very fast time-to-value with the data they’re replicating into the Context Store. It is also a way for us to get into the discipline that I think every company needs to adopt — don’t just feed a playbook to an agent. Actually build a workflow, and make sure that you can iterate as models are evolving and as you’re discovering things that are working and things that are not. It’s really taking an initiative to help our customers on how to start creating agentic workflows. It’s in preview — we also want to make sure the market and the overlap with what type of automation you can do on Claude versus Airbyte is balanced. But for us, it’s a very simple way for people to just get started and feel comfortable building agentic systems on top of the data that they own.”
Agent Operations: Pricing That Aligns With Agent Value
Airbyte has introduced a new pricing unit — agent operations — designed to align cost with the value agents actually deliver. Every successful retrieval, read, or write action against the Context Store or a connected source system counts as one agent operation. Unsuccessful or irrelevant fetches are not charged. The model is designed to scale proportionally with the real-world utility agents are generating.
Q: You’re metering consumption in something you call agent operations — what is that unit and how should teams think about cost as they scale?
Michel Tricot: “Agent operation is really tied very closely with the agentic data loop we talked about — the ability to discover. The ability to retrieve, to search for specific information on the Context Store, is something that will be counted as an agent operation. We only charge for the ones that are successful and return value. Then you have the ability to read directly into the source system, because sometimes you need data that might not be in the context. Finally, you have the ability to act — writing data — and every action you make toward a downstream system is something we manage. As the product evolves, we’re going to add more and more intelligence to make sure that agents make fewer and fewer mistakes. The more successful people are, the more they rely on the data we provide — well, the more they consume on the product.”
Michel Tricot: “The piece around Automation — same thing. We’re going to be charging with some formula related to the number of tokens that are being read and that are being written, especially because it’s a data product. There will be a lot of input tokens taken into account because we provide the data, and that’s something that is reflected in the agent operation as well.”
Airbyte’s Identity Shift: From Data Movement to Agentic Data Platform
Since its 2021 open-source launch, Airbyte has moved data into data warehouses. That was the assumed destination for every connector in the catalog. Starting in early 2024, something changed: users began pointing connectors at vector databases, plain file storage, and embedding pipelines. When Airbyte talked to those users, they discovered an entirely new generation of data processing pipelines — pipelines built not for analytics, but for AI. That signal was the origin of the Airbyte Agents launch.
Q: You started as open-source data integration — now you’re describing yourself as an agentic data platform. How does the open-source community and your existing enterprise customers fit into what you’re building?
Michel Tricot: “What got us to start investigating that extension of the product is that we saw a change in pattern in how people are using the open source product today. Since we first released Airbyte in 2021, every piece of data was going into a data warehouse. What we’ve seen starting at the beginning of 2024 is that people were asking us for different types of destinations — yes, vector databases, but also people asking us to write plain files directly into object storage. When we talked to these community members about what they were doing, they said, ‘Yeah, we’re using Airbyte for something other than analytics.’ Airbyte is a data movement platform, so for me it doesn’t really matter where the data is going — but I need to understand: if I see a change in pattern, what is the new use case being worked on? And what they were building was a new generation of data processing pipeline where they would pull unstructured data, pull structured data, do embedding on that data, feed it into a vector database, try to inject structured information into unstructured information to make sure that retrieval with RAG architecture would work. To me, that was a very strong signal that analytics is here, it works, but there is an extension — a new type of consumer. I want to make sure that as a data movement company, we can actually address these new emerging use cases.”
From Months to a Week: A Real Customer Case
One of the most concrete proof points Tricot shares is a marketing platform customer that needed to build an agent helping their customers manage campaigns across Facebook, TikTok, HubSpot, and internal databases. They started building the integrations themselves — and quickly realized their entire roadmap would become integration work. Airbyte’s Context Store replaced that entire integration backlog, giving their end customers a simple credential widget and delivering live, indexed data in minutes.
Q: Talk about what the actual blocker was that Airbyte Agents removed for customers, and how someone went from months of work to production in a week.
Michel Tricot: “The first thing that saved them a lot of time was that we are experts at connecting to external systems — that is really where we excel. They had a marketing product, and what they needed was to build an agent to help their customers build campaigns, monitor campaigns, create visual assets, and so on. In order to do that, they had to fetch information across Facebook, TikTok, HubSpot, and their own databases — and they actually started to build it themselves. Then they realized: we are done with one integration, now we need to do the second, the third, the fourth — plus we also need to interact with these APIs. For them, the whole roadmap was going to be about connecting to systems.”
Michel Tricot: “What we did with them was give them a very simple widget. They could provide it to their customers, customers input their credentials across all the systems they use, and we give them access to all this data — now available in the Context Store — in just a matter of minutes. The moment the customer has included their credentials, you wait a few minutes and the data is available. Now their agent can start operating on it, tuning campaigns, creating new campaigns. And that is something that really takes a lot of time, especially for marketing where you have potentially massive budgets. They were able to go to market in barely a week with the Context Store that we’re providing to them.”
Does the Context Store Replace Existing RAG Pipelines and Vector Databases?
Many enterprises already have RAG pipelines and vector databases in place. Tricot’s position is nuanced: the Context Store is partly a replacement for naive RAG approaches that rely solely on unstructured data indexing, but it is primarily an upgrade — adding entity resolution, cross-system structured data linkage, and permission enforcement that standalone RAG pipelines cannot provide.
Q: Enterprises already have vector databases and RAG pipelines. Where does Airbyte Agents sit relative to those — are you complementing them or replacing them?
Michel Tricot: “There is a replacement aspect to it, because a RAG pipeline is great, but it rarely solves the problem of cross-system retrieval. RAG pipelines — people at the time thought they were the magic bullet that would make LLMs a lot smarter. But what people realized is that the moment they start doing something RAG-related, they go to index a sales call. So now you index a sales call. What do you have at your disposal? You have access to potentially the title of the meeting — better not to have that meeting called ‘call with Swapnil,’ because at that point I don’t know who Swapnil is. I don’t know what company he’s part of. You might not even talk about the company within the call. So what people had to do when building this pipeline was, for every single chunk of unstructured data they were indexing, they had to put a little header saying, ‘Swapnil is part of Company A, here’s some information about Swapnil,’ in order to make the search relevant. That was basically working around the fact that indexing just on unstructured data and trying to structure data inside of it to make search more relevant is a massive waste of resources and energy, and it’s very imperfect.”
Michel Tricot: “Whereas with the Context Store, you get basically the best of both worlds — the structured data is properly indexed and it’s actually a link for you to get to the proper structured data. And the agents, if they don’t have the right information from that entity, are going to look at the attached entities. That helps the agent actually find what is the most relevant call if I’m talking to Swapnil about pricing and about an issue he was encountering with his own customer ABC. Rock pipelines are good, but if you want an agent to actually be efficient, you need to have more than just chunks of unstructured data. You need to be able to attach it to real structured data that is coming from the systems you work on every single day.”
Getting Started: Pricing, Onboarding, and the Path to Enterprise
Airbyte Agents is designed for a bottom-up adoption motion — individual users get to value first, then expand to team plans, then to enterprise. Tricot emphasizes that the product has been built specifically to minimize time-to-value for a single user, before scaling outward. Enterprise pricing is built on top of Airbyte’s existing enterprise infrastructure.
Q: How can teams get started, and how involved is Airbyte in the onboarding process?
Michel Tricot: “Culturally at Airbyte, we have always had a very good relationship with end users, and we’ve built more of an enterprise motion over the past few years. For this product, especially with how companies are adopting AI, it is part of the end user to figure out what is a good solution for the problem at hand. My team has been asked and tasked to identify good AI solutions. They’ll test it and make a suggestion about this particular product. So for Airbyte Agents, everything has been tailored for one user to be able to get to value as fast as possible, and then promoting to more team-type plans. For companies that need a lot more seats, connections, or types of data, they have access to more enterprise pricing built on top of all the enterprise features we’ve already built within Airbyte. But the goal for us right now is to get as many end users to feel the value of Airbyte and what they get from getting the Context Store connected to their agents or their chatbot.”
