Research & Deep Dives
Research papers, reports, scenarios, and technical findings translated into practical implications for builders and operators.
Open Research & Deep DivesA survey argues that agent skills are becoming the packaging layer for procedural AI work, making governance, permissions, and verification unavoidable. It maps the controls needed before reusable instructions can be trusted across teams and organizations.
A survey argues that agent quality is a system property spanning models, memory, tools, planners, verifiers, permissions, traces, and evaluation. It offers a practical architecture for diagnosing failures without blaming the base model by default.
A NeurIPS dataset paper finds that multi-agent LLM systems fail through role confusion, broken handoffs, and weak verification, not only weak models. Its taxonomy turns coordination breakdowns into observable failure modes teams can test and repair.
A paper argues that agent skills need a build-time optimization pass because many reusable instruction files waste context and make agents worse. Its proposed compiler trims redundancy while preserving the instructions that actually improve task performance.
StoryScope suggests that AI fiction can be detected from narrative decisions, not only surface style, including tidy plots, explicit themes, and reduced structural variety. The findings suggest provenance detection should examine story structure alongside lexical fingerprints.
A NeurIPS position paper argues that LLM social simulations need traceable belief models, not merely fluent personas and plausible survey answers. Explicit belief state, update rules, and validation make simulated populations more scientifically useful.
A systematization paper argues that reusable agent skills are becoming the procedural layer of agent systems, with meaningful performance gains and supply-chain risk. The survey maps benefits alongside provenance, permissions, poisoning, and dependency management.
A preprint argues that agent reliability is governed by the runtime harness around the model: execution, tools, context, state, lifecycle policy, and evaluation. The framework gives teams a way to reason about reliability beyond benchmark scores.