---

---

22nd May 2026

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

• SkillOpt: introduces a systematic text-space optimization framework for agent skills that treats the skill document as a trainable external state for a frozen Target Model, utilizing an Optimizer Model to propose bounded edits based on Rollout Batch evidence.
• The framework employs a Validation Gate to ensure only performance-improving edits are accepted, while a Rejected-Edit Buffer and epoch-wise Slow-Update Field provide negative feedback and long-horizon stability without requiring model-weight updates.
• By decoupling the execution harness from the optimization loop, SkillOpt enables the creation of compact, reusable, and inspectable skill artifacts that transfer effectively across different model scales, execution environments, and benchmarks.

---

From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills

• SkillLens: introduces a utility-grounded evaluation framework that systematically analyzes the full lifecycle of model-generated agent skills across experience generation, skill extraction, and skill consumption.
• The framework identifies that skill utility is not determined by textual plausibility but by concrete failure mechanism encoding and actionable specificity, which are operationalized into a meta-skill to improve extraction.
• Experimental results across five diverse domains demonstrate that while model-generated skills are generally beneficial, they exhibit non-trivial negative transfer, necessitating a principled, utility-grounded approach to skill development.

---

CHRONOS: Temporally-Aware Multi-Agent Coordination for Evolving Data Marketplaces

• CHRONOS: introduces a three-layer architecture for temporal KG marketplaces that integrates T-LEGEND, Event-Conditioned MPV, Temporal Coordinator, and a Fixed-Dimension DP Pipeline to provide end-to-end temporal guarantees.
• The framework utilizes neural-ODE temporal decay for index freshness, BOCPD-conditioned Shapley valuation for fair attribution, and an EXP3-IX bandit agent for coordinated DP budget manage