AI Agents Make Unified APIs Obsolete
Ayan Barua, CEO of Ampersand, on infra for AI agent integrations
The strategic shift is that value is moving away from flattening software into a lowest common denominator schema, and toward infrastructure that helps agents work directly with each customer’s real, messy system. Unified APIs were useful when the goal was to get a human developer to a basic integration fast. In an agent world, the harder problem is reading tenant specific fields, handling permissions, retries, and writing safely across many systems in real time.
-
Unified APIs still fit narrow, repeatable jobs, especially analytics style use cases in categories like HRIS, where a product mainly needs a standard employee or payroll record. But they break once customers need custom objects, extra fields, or tenant specific workflows, which is where enterprise integrations usually become painful.
-
That is why infrastructure players are converging on a deeper stack. Ampersand is adding MCP and an AI SDK on top of native integrations. Vessel pairs a unified layer with an action API for tool specific edge cases. The common pattern is not one schema to rule them all, but a shallow default with a path to full system depth.
-
What survives is the abstraction around operational headaches, not the abstraction around data itself. Agents still need authentication, token refresh, tenancy isolation, rate limit handling, observability, and safe read and write controls. MCP helps agents talk to software, but it does not remove the need for that orchestration layer.
The next few years should turn integrations into agent infrastructure, where the winning products are the ones that let software act across many systems with fresh context and enterprise grade controls. That favors platforms that can expose data deeply, move it in real time, and keep humans out of field mapping and dashboard clicking for most routine work.