SourcesPapersWweb·medium signalXBlueskyLinkedInCopy link**SkillRL: Recursive Skill-Augmented Reinforcement Learning** — [arXiv](https://arxiv.org/abs/2602.08234) (medium). Agents learn reusable skills organized hierarchically. 10-20% token compression, 15.3% improvement on ALFWorld/WebShop. Most practical paper on making agents actually improve from experience.↳ Follow the threadPolicy dependency / Stack layerDistilling 1,000 ML repos into 5,000 verified skills lifted an agent 134% on MLE-bench with the model and budget held fixedarXiv 2609.02749Stack layer / ContrastBelief-Calibrated Optimization makes the coding-agent-as-optimizer state its hypothesis before editingarXivPolicy dependency / Stack layerCo-evolving the harness alongside the policy cut AgentDojo attack success 3x while raising benign utilityarXiv 2609.02786Stack layer / Update threadMulti-agent systems can learn continually through skill libraries with credit assignment, not reflection memoriesarXiv 2609.02094Stack layer / Update threadThe Near-Optimal SFT-to-RL Budget Split Is Wide and Transfers From Small Proxy Models to Large TargetsarXiv 2609.01573Policy dependency / Stack layerCROSS-CATEGORY: Four Unrelated Vendors Shipped Agent Authorization Control Planes Inside 48 HoursJetStream (corroborated by Genesys Xperience 2026 coverage, aiagentstore.ai and Hacker News Show HN)Policy dependency / Stack layerFour advisories land on Databricks' Omnigent meta-harness, one critical, all reported by an autonomous security agentGitHub Security AdvisoriesStack layer / ContrastSMELT Loops the Middle Half of an MoE Transformer Twice and Saves 6.8-18% of Training FLOPsarXiv 2609.01343