![](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/05\/airbyte-logo.jpeg\"){kind=logo}
\n Airbyte\u2019s CEO argues AI agents fail in production because they waste tokens assembling context at runtime. His fix: pre-index data into a unified Context Store.<\/p>\n
\t\t\t\t
\n Airbyte\u2019s CEO argues AI agents fail in production because they waste tokens assembling context at runtime. His fix: pre-index data into a unified Context Store.<\/p>\n
Airbyte,<\/a> the data movement platform used by thousands of companies to move billions of records a day, has launched Airbyte Agents \u2014 a new product built around an infrastructure pattern called the Context Store that the company says aims to resolve the root cause of AI agent failures in production environments.<\/p>\n The Five-API-Call Trap Also Read:\u00a0AIThority Interview With Rohit Agarwal, Founder & CEO of Portkey<\/a><\/p>\n Each sequential call: Tricot calls this \u201cruntime context assembly\u201d and says it is the central failure mode of enterprise agents today, one that neither better prompts nor orchestration frameworks can fix. He also notes that Model Context Protocol (MCP) does not resolve it: \u201cA stack of MCPs still forces your agent to hunt through systems one at a time, burn tokens on raw data, and miss how any of it connects. MCP gives you access. It does not give you understanding.\u201d<\/p>\n The Context Store: Moving Work Upstream \u201cYour business becomes a living model,\u201d Tricot says. \u201cEvery entity, from every system, all in one place. The whole picture stays fresh in the background. Any capable agent can reason across all of it in a single query.\u201d<\/p>\n Tricot says agents using the Context Store make 40% fewer tool calls and consume up to 80% fewer tokens, reducing latency and cost at scale.<\/p>\n From Pattern to Product \u201cThe data industry learned this lesson a decade ago,\u201d he says. \u201cThe hardest part of any data problem is never the compute. It\u2019s getting the right data, in the right shape, to the right place, at the right time.\u201d<\/p>\n Also Read:\u00a0\u200b\u200b<\/a>AI-Driven Risk Intelligence: How FIs Are Predicting Systemic Shocks<\/a><\/p>\n
Michel Tricot, CEO and co-founder of Airbyte, argues that the failure is architectural, not computational. When an agent answers a business question \u2014 such as the status of a client renewal \u2014 it typically makes five or six sequential application programming interface (API) calls across Salesforce, Zendesk, HubSpot, Slack, and contract tools before it can reason over the result.
\u201cThe models are smart enough,\u201d says Tricot. \u201cThe problem is that we keep handing brilliant reasoning engines terrible data and expecting good results.\u201d<\/p>\n
Adds latency
Burns tokens on raw, developer-formatted data the model often does not need
Returns stale or contradictory results when systems were updated at different times
Fails if any single API is rate-limited, paginated, or down
Why New Protocols Don\u2019t Fix It<\/p>\n
The Context Store inverts this approach. Rather than assembling context at query time, it continuously replicates and pre-indexes data from all business systems into a unified layer where records about the same entity \u2014 a customer in Salesforce, their tickets in Zendesk, their contract in billing \u2014 are already matched and linked. When an agent queries this store, it makes one call and receives clean, structured data in under a second.<\/p>\n
Airbyte Agents ships the Context Store alongside three interfaces: a software development kit (SDK) for engineers, an MCP integration for AI clients such as Claude and ChatGPT, and a no-code builder for business teams. Tricot says the underlying pattern is model-agnostic and expects it to become standard data infrastructure.<\/p>\n