Cognaptus Insights

Terra Nova shows why serious agent evaluation must test coupled strategy, uncertainty, cooperation, and long-horizon trade-offs rather than another tidy task list.

A mechanism-first reading of TIM, a multi-agent LLM framework that turns opaque DeFi transactions into evidence-ranked intent labels without pretending to read private motives.

A field experiment in AI-authored and AI-reviewed science shows that research agents are useful only when wrapped in disclosure, verification, and human judgment.

Memory-R1 shows why durable AI agents need learned memory operations, not just bigger context windows or more enthusiastic vector search.

A mechanism-first reading of Octopus, a multimodal agent framework that treats reasoning as capability orchestration rather than a bigger-model contest.

A mechanism-first analysis of how rate–distortion theory and fused Gromov-Wasserstein alignment can make educational knowledge graphs more useful, not merely larger.

A mechanism-first look at how heterogeneous multi-agent reinforcement learning could turn distribution-grid restoration into faster, constraint-aware decision support.

A close reading of why LLM-generated optimisation models can look correct, compile occasionally, and still misunderstand the problem hiding in plain sight.

SkillGen shows why the next gain in LLM agents may come from reusable procedural skills, not longer prompts or larger models.

How calibrated symbolic uncertainty helps robots decide when to act, when to look again, and when confidence becomes expensive.