Airbyte Agents fly in to fix the data breaks
Open data movement platform creator Airbyte has this month detailed Airbyte Agents, a context layer designed to gives production-grade agents access to a unified, search-optimised index of a company’s data.
That index is replicated and ready to query before the agents run.
The company tells us that most agent failures in production are not model failures, they are “data failures” (a term that may have been coined in a business publication by deep learning pioneer Andrew Ng.
Agents built on runtime API orchestration chain together five or six calls across disconnected systems to answer a single question, burning tokens, adding latency, and frequently returning stale or contradictory results
Airbyte claims its new Agents service solves this at the data layer rather than the orchestration layer.
Context is core
At the core of Airbyte Agents is the Context Store, a replicated, search-optimised index that unifies a company’s data across systems before the agent ever runs. Customer records from Salesforce, tickets from Zendesk, issues from Jira, and conversations from Slack are brought together into a single queryable index with history and state preserved. The work of assembling context happens in advance, not at query time, so agents query the Context Store instead of chasing live APIs.
That typically collapses five or six calls to one or two and dramatically reduces token consumption.
“Airbyte’s Agent Engine has massively accelerated our roadmap. What we thought would take 6-plus months, we were testing in the first week of the beta program,” said Nate Chambers, chief product officer, ORCA Analytics. “They’re shipping everything we need for agentic workflows, and launching new connections faster than we can build them into our product. If you’re building an AI product, you can stop rolling your own pipelines and start shipping.”
The platform is available through the Model Context Protocol (MCP), which works inside Claude, ChatGPT, Cursor, and any MCP-compatible client, and through a native SDK for teams building custom agents from the ground up.
“Most agent projects stall for the same reason: The model is fine, the data is a mess. Five disconnected systems, inconsistent entities, no shared state,” said Michel Tricot, co-founder and CEO of Airbyte. “Airbyte Agents gives every agent a unified view of the business, replicated and ready to query. That is what separates an agent that can do the work from one that just talks about it.”
Airbyte Agents is available two ways:
- Airbyte MCP: Connect data sources to Airbyte once, then build and run agents inside Claude, ChatGPT, Cursor, or any MCP-compatible client. No code required, and the same governed access to the Context Store that the SDK provides.
- Agent SDK: For teams building custom agents and applications directly against the Context Store, with full programmatic control over retrieval, permissions, and state.
The platform launches with 50 connectors that populate the Context Store, covering the systems most central to enterprise operations, including Salesforce, HubSpot, Zendesk, Jira, and Slack. Airbyte’s catalogue of 600+ connectors will be available in the Context Store in the months ahead.
OAuth-based authentication
A growing share of connectors also support write actions, letting agents update records, create tickets, and post messages in the systems of record. All connectors support OAuth-based authentication and row-level permissions, so agents only see what the invoking user is allowed to see.
Automations, a visual interface for building and running agents directly inside Airbyte, is also available in research preview. Built on the same Context Store as Airbyte Agents, it lets teams compose agentic workflows across connected systems without code, and will graduate to general availability in a later release.