+the observation that LLMs fail in predictable ways — sycophancy, context drift, hallucinated confidence, inconsistency across conversation turns — and that these failures are not just model problems but also prompt engineering and scaffolding problems. the idea is to systematically study and mitigate these failure modes through a combination of better prompting patterns, structured context management, and behavioral testing frameworks.

+

+one concrete direction: building a suite of tests that probe specific behavioral failure modes (does the model change its answer when the user pushes back? does it maintain consistency over a long conversation? does it respect negative constraints?). another direction: studying what kinds of AGENTS.md / system prompt patterns produce reliably better behavior, which overlaps with [[agents-md-research|AGENTS.md research]]. the meta-insight is that much of what people attribute to "bad AI" is actually addressable at the prompt and scaffolding layer without waiting for better base models.

+

+this connects directly to [[context-window-optimizer|context window optimizer]] which attacks one specific failure mode (context degradation over long sessions), and to [[spec-driven-dev|spec-driven dev kit]] which applies structured prompting to coding tasks. [[cognitive-foom|cognitive foom]] is the more ambitious version — recursive self-improvement infrastructure for agents. [[llm-physical-intuition|LLM physical intuition]] is a related empirical research question about what spatial/physical reasoning current models do or don't have.