Michel Tricot, CEO and Co-Founder, Airbyte

Can you introduce yourself and your role at Airbyte?

My name is Michel Tricot, and I’m the CEO and Co-Founder of Airbyte.

Can you tell us more about Airbyte as a company?

Airbyte makes data available and actionable to everyone, everywhere. Jean Lafleur and I founded the company in 2020 through Y Combinator (W20).

Today, Airbyte is the leading open-source data movement platform, supporting 600+ connectors across data replication and AI agent use cases. 

What did Airbyte recently launch, and what problem does it solve?

We’ve released Agent Engine into the public beta at app.airbyte.ai. It’s a unified infrastructure layer that gives AI agents fast, structured access to business data, addressing the context problem that has stalled agent deployments in production. 

For the past two years, the AI conversation has focused on models — how powerful they are, how fast they’re improving, and which one to use. But agents have not realized the same progress. Today’s leading models from OpenAI, Google, and Anthropic can reason, plan, and use tools. The limiting factor is no longer intelligence. The real bottleneck is how agents assemble context.

Why are AI agents hitting a wall in production?

While application programming interface (API) and Model Context Protocol (MCP) access to business tools is widely available, it doesn’t give agents a coherent, unified, 360-degree view of the business. Instead, they assemble context in real time — pulling fragments from different systems, one API call at a time.

That approach introduces the same problems, over and over:

• Latency: Every system adds delay.
• Token waste: Raw data overwhelms the model.
• Fragmentation: Multiple versions of the same “truth”
• Brittleness: Failures in auth, rate limits, or APIs

The result is predictable: the demo works, the production system breaks.

How do you define the “context problem” in AI agents?

Context is the key to making AI agents work. Without context, AI agents cannot reason and are doomed to failure.

Typical agent systems are built on a runtime API orchestration chain that involves five or six API calls across disconnected systems to answer a single question — burning tokens, adding latency, and returning stale or contradictory results.

What does “one API call” mean in the context of Agent Engine?

AI agents need real-time access to data sources — such as Salesforce, Zendesk, and others — that include context such as awareness of relationships, history, and state in one API call.

That means no more latency issues, no token waste, and a system that scales.

How does Agent Engine connect to business systems?

Agent Engine launches with an initial 20+ connectors — including Salesforce, HubSpot, Gong, GitHub, and Linear — and connects agents to business systems in approximately 10 lines of code.

A fully-managed authentication module supporting Open Authorization (OAuth) handles credentials, so teams can focus on agent logic instead of integration plumbing.

What is the Context Store, and how does it work?

At the core of Agent Engine is the Context Store, a governed data layer that pre-indexes key fields from connected sources. Agents can search across it in half a second or less.

It enables agents to have fast, structured access to business data with a coherent, unified view, instead of assembling fragments across systems in real time.

What makes this approach different from typical agent systems?

Typical agent systems rely on a runtime API orchestration chain across disconnected systems, requiring multiple API calls to answer a single question. This approach burns tokens, adds latency, and returns stale or contradictory results.

With Agent Engine, instead of assembling context in real time through multiple API calls, we provide a unified infrastructure layer with a governed data layer where context is pre-indexed and available. This eliminates latency, reduces token waste, avoids fragmentation, and minimizes brittleness caused by failures in auth, rate limits, or APIs.

We’re effectively turning multiple API calls across disconnected systems into a single, reliable query.