Raw LLM Responses

I probably have similar usage patterns as you do (Claude, Codex, etc.), and yes, agents do the darndest things, but I disagree with your characterization. Any system like this will have a percentage distribution where it does what we want, somewhat does we want, or completely ignores us, hallucinates, etc. These percentages have been continuously shifting towards more useful tasks being achievable. But as humans, we'll always take the pathological case ("the agent ate my homework") and call the system a glorified pattern-matcher for making such dumb mistakes a human would never make. That's an extremely reductionist (typically human) view of intelligence, which is a high-dimensional set of capabilities. In reality, the intelligence of the models has been continuously increasing across many dimensions. But humans want to see some kind of magical "AGI" threshold being crossed before they concede intelligence. Case in point, here's NB Pro illustrating that concept with two quick prompts, something that would not have been possible a couple of months ago: https://preview.redd.it/zdb55xgka96g1.png?width=1024&format=png&auto=webp&s=e66467cb7067c11a28de35a78dde85c6464f484b