Tagged observability · show all articles
Tool protocols standardize how a model asks a server to act. They say far less about what happens when a schema changes underfoot, a timeout fires before a write commits, a token expires mid-task, or a session drops — production finds every one of those gaps.
An agent architecture is five engineering decisions wired into one harness — permission scoping, state persistence, checkpoint gates, retries, and instrumentation — each documented by a real SDK, each with a demonstrated production failure mode.
A production retrieval system is not one architecture but six coupled engineering decisions — chunking, hybrid weighting, reranking, access control, index freshness, and what gets watched once it is live.
The dangerous failures are not the ones that throw an exception. They are the ones where the system keeps returning well-formed output that is quietly wrong, and nothing upstream notices.