Agent-Centric Integration Infrastructure
Ayan Barua, CEO of Ampersand, on infra for AI agent integrations
This shift changes the job of integration software from serving programmers who know exactly which endpoint to call, to serving agents that need safe, legible ways to explore, read, and act across many systems. MCP matters because it gives agents one standard way to ask for context and take actions, while companies like Ampersand sit above that protocol to handle the hard production work, like auth, field mapping, retries, rate limits, governance, and tenant isolation.
-
In the old API world, a developer wrote custom code against each app. In the MCP world, the agent can talk through one common interface to many tools. That makes the integration surface less about endpoint memorization, and more about exposing clean tool definitions the model can reliably use.
-
This does not remove developers, it changes where value sits. The protocol handles the common language, but the winning infra layer still needs to manage real enterprise messiness, like one customer having 100 Salesforce tenants, custom objects, shared rate limits, and strict privacy boundaries between accounts.
-
The same pattern is showing up across the market. Merge now offers an MCP server so one agent can access hundreds of tools through a single server, and Finch says its MCP server lets LLMs use payroll and HR APIs through natural language instead of hand written API calls.
Over time, business software is likely to look less like people clicking through dozens of dashboards, and more like agents sitting on top of systems of record and pulling live context when needed. That pushes integration vendors toward becoming the control plane for agent actions, where reliability, security, and real-time context matter more than the developer facing API alone.