Airbyte CEO Michel Tricot published a perspective in his latest blog post arguing that treating the model as a stateless utility behind a routing layer and owning it rather than renting it is the architecture we've bet on.

Founded in 2020, Airbyte is the context layer for production-grade AI agents. AI agents fail in production for one reason: they can't see the business. They make scattered API calls at runtime, burn tokens reconciling fragmented data, and break under real workloads. Airbyte solves this by giving agents unified, permission-aware access to the operational data scattered across the tools companies actually run on (CRM, billing, support, product, internal systems) through a hybrid architecture that combines large-scale replication (for cross-system search and discovery) with real-time fetching (for fresh operational state). It's the combination production agents actually need, built on the connector footprint we've hardened over six years. We've raised $181M from Benchmark, Accel, Altimeter, Coatue, Y Combinator, and others. Today, 25,000+ companies sync data with Airbyte's 600+ connectors. Open source remains core to how we build, because the data foundation under your AI agents is too important to be a black box.