I let the agent test its own model upgrade instead of trusting the release notes. It found 3 things throttling itself

I swap the model under my local agent fairly often. Usually I read the reviews, flip the config, and hope. This time I tried something different, Im away on holiday but took my dev righ with me, on hotel WIFI 😄 I upgraded the model additively (new one on a separate port, old one still a keystroke away), then pointed the agent now running the new model at the reviews I saw from YouTube as well as HuggingFace model cars and told it to grade its own upgrade. It didn't summarize anything. It spun up a throwaway test server on a third port so it wouldn't disturb my session, fired controlled probes at itself, watched the GPU pin at 92%, measured its own decode speed firsthand instead of trusting a reviewer's number, and read its own weight index. It did a stellar job in my mind, i did have a off ramp just in case tho... Do you let your agents verify their own tooling/config, or do you keep that human-in-the-loop? Curious where people draw that line.