Airbyte launches Agents with a unified data layer

Airbyte has launched Airbyte Agents, a new data layer meant to help AI agents find information and take actions across operational systems more reliably. Michel Tricot, the company’s co-founder and CEO, announced the product in a Show HN post and said it is built on top of Airbyte’s existing connector infrastructure.

According to Tricot, the goal is to solve a problem that shows up when agents move from demos into real workflows. Agents often need to work across tools such as Slack, Salesforce, Linear, Gong, Zendesk, and GitHub, but each system brings its own authentication, pagination, filters, schema differences, and entity-matching issues.

Tricot said most MCPs, or Model Context Protocol implementations, do not solve that problem by themselves. He described many of them as thin wrappers over APIs, which means agents still have to work with the underlying API’s limits and often need to know what to query before they can reason about the task.

Airbyte Agents is positioned as a context layer between an agent and the underlying systems. The core component is something the company calls Context Store, which Tricot described as a data index optimized for agentic search and populated by Airbyte’s replication connectors.

The idea is to let an agent discover relevant data in a structured way before making decisions. At the same time, Airbyte says agents can still read from and write to the source system directly when that is needed.

Tricot said the product was motivated by a specific agent trace the team encountered while migrating an agent to Airbyte’s new SDK. The task was to answer a business question: “which customers are at risk of leaving this quarter?” According to Tricot, the agent took 47 steps, most of them API calls, as it tried to find accounts, map them to customers, and search for support tickets.

He said the final answer sounded reasonable but was wrong, and the process was also very slow. That example pushed the team to build a better way to assemble context before the agent starts reasoning.

Airbyte said the new system performs especially well on tasks that require joining data across multiple tools. Examples given by the company included, “Show me all enterprise deals closing this month with open support tickets,” and “Find every support ticket that doesn’t have a Github issue opened.”

To evaluate the product, Tricot said he built a benchmark harness over a weekend and compared the Airbyte Agent MCP against vendor MCPs directly. He used token consumption as the measure, saying it is a useful proxy for how well an agent is working.

According to the results he shared, Airbyte Agents used up to 80% fewer tokens than Gong’s own MCP, up to 90% fewer than Zendesk’s, up to 75% fewer than Linear’s, and up to 16% fewer than Salesforce’s. Tricot noted that Salesforce’s own SOQL query language already performs well for this use case.

Airbyte also acknowledged the limits of the benchmark. Tricot said the company built the system and the test, so there is obvious bias. To make the evaluation public, Airbyte published the benchmark harness on GitHub at airbytehq/airbyte-agents-benchmarks.

The company said the product is still early and some parts are rough, but it wanted to share the release with the community quickly. Airbyte is also asking builders whether they index data ahead of time or let agents call APIs live, and how they match entities across systems.