Topic
The design patterns and tradeoffs in building software on top of large language models, covering state management, prompt design, and tool integration.
A dev.to post lays out a four-layer recommender and a 100 millisecond p99 budget with no slot for a generative call in the hot path. The figures are its author's own allocation, not a measurement from a running system.
Reality
Statewave's case for a separate agent memory layer rests on three defaults in a RAG stack: similarity-only ranking, append-only chunks, and no compaction. The post asserts all three failure modes and measures none of them.
Reality
Anthropic's December 2024 guidance and three agent benchmarks converge on an awkward result, because the workloads whose steps cannot be enumerated in advance are also the ones where measured agent completion is lowest.
Reality
A vendor blog argues the hole is architectural: one token stream, no command/data channel. Its remedy list runs to five controls, and not one of them is a prompt.
Reality
A dev.to post argues prompt hardening is the weakest defence against injection, not the strongest. If your security depends on the model choosing to obey, you have a suggestion.
Reality
A dev.to walkthrough argues that routing every job through one 'strongest model' helper hides the economics. The cheap fix is declaring task class and requirements before the request leaves your code.
Reality
A procurement team swapped a trained classifier for an LLM to pick one cost centre out of thousands. The talk transcript reads as a list of the boundaries you have to rebuild by hand.
Reality