One of my favorite Slack feeds at Sanity is a steady stream of agent suggestions in response to actions they tried to take on our platform via our model-context protocol (MCP) server. Agents are happy to tell you exactly how they tried to do something, why they think it failed, and what they need from you. Agent feedback makes it easy to knock out tiny improvements daily (it often comes with detailed error repro steps).
The product almost hardens itself if you pipe that feedback somewhere that's easy to query. Example: I built a bridge solution in Sanity MCP for a tricky problem involving passing token-inefficient data outside the model's context window and saw a month-over-month doubling of "I can't use <what-you-built>, I need ___ because ___" straight from agents who tried and failed.
People are busy: they may silently work around the sharp edges of what you’ve built and leave it to you to discover what’s working and what’s not. Let their agents help you help them by asking for feedback when things go wrong.