How should an internal app support both human users and AI agents?

Don’t forget the browser layer if agents may operate the application through the UI. Give important controls stable, semantic identifiers and accessible labels so automation can reliably find them. That said, direct API access is generally more robust than having an agent click through the web interface, so treat UI automation as a fallback rather than the primary integration.

It can help to think of AI as another client with different constraints, not as a completely separate type of user. Design the permissions, audit logs, confirmations for risky actions, rate limits, and idempotency carefully. Then add agent-friendly documentation and a small set of high-value tools based on real workflows instead of exposing every internal endpoint at once.

The main difference is usually not the core functionality but how much data and ambiguity each interface exposes. A response with hundreds of records and dozens of fields may work for a human table view but consume an agent’s entire context window. Provide focused queries, useful filters, pagination, concise response shapes, explicit action names, and APIs that make the next valid step obvious.

Your basic approach is sound: make the underlying application and API reliable first, then expose the same capabilities to the web client and to AI tools. An MCP server can be useful as an adapter for agent tooling, but it shouldn’t replace a clean domain API. Start with clear operations, predictable errors, strong authentication and authorization, good documentation, and safe defaults.