A Better Newspaper

A set of October 2026 papers explore how LLM agents can accumulate, organize, and refine reusable 'skills' from past experience to improve performance over time without retraining base model weights. ## Key Approaches - **Cognitive Differentiation** proposes organizing agent skills by functional heterogeneity rather than treating them as monolithic artifacts, drawing on cognitive-science concepts of differentiation to improve knowledge maintenance and use (arXiv:2610.23599, October 20). - **Darwin-Claw** introduces environment-adaptive experience consolidation for recursively self-evolving agents, aiming to avoid indiscriminate injection of raw trajectories into memory that obscures behavioral patterns (arXiv:2610.23574, October 20). - **SkillBrowser** addresses 'Selective Skill Exposure,' balancing the need to expose relevant procedural skill content against the risk of context pollution from overly broad disclosure or static retrieval (arXiv:2610.23554, October 20). - **Harness-Delta Attribution (HDA)** proposes an evaluation protocol to determine how much of the performance gain from harness evolution (prompts, skills, tools, orchestration) reflects genuine generalizable improvement versus benchmark overfitting or extra inference compute (arXiv:2610.22466, October 20). ## Why It Matters As agentic AI systems move toward persistent, self-improving deployments, the mechanisms by which they store and reuse 'skills' will shape both their capability ceiling and their failure modes. Rigorous attribution methods like HDA are important for enterprises evaluating whether claimed agent improvements will generalize to production use, rather than reflecting narrow benchmark tuning.