someshutkar

I can see why this is becoming a bigger topic.

How are you handling changes over time? Business rules, permissions, and approval workflows tend to evolve, so I'm curious whether governance policies are versioned independently from the agents themselves.

I don't think model drift is the only issue. In production, I've seen more failures caused by changes around the model, such as APIs, retrieval quality, data pipelines, permissions, and business logic. Even with a stable model, the surrounding system continues to evolve. That's why observability and continuous evaluation at the application layer remain essential.

I like the idea of separating access control from memory itself. One question I had while reading this: how do you prevent the agent from pulling in too much context? In production systems, we've found that deciding what not to retrieve can be just as important as maximizing recall.

Nice to see the focus on production rather than just prototyping. I'm particularly interested in the Context7/MCP integration—did you evaluate other approaches before settling on this architecture?