Did Codex Reset
GitHub

SkillLift paper reports 40–70% lower token cost for evolving agent skills

DAIR.AI

DAIR.AI highlighted a paper on SkillLift, a method for evolving agent skill prompts without running a full agent rollout to score every candidate revision. SkillLift trains a rubric to agree with real outcomes on which of two skills is better, on the grounds that ranking needs fewer oracle runs than predicting each score.

An inner loop revises skills against a frozen rubric at no rollout cost. An outer loop spends a few real rollouts to re-align the rubric by rank correlation. The post says rollout cost is why skill self-evolution usually only patches observed failures.

On SkillsBench and WildClawBench (147 tasks) with three models, SkillLift reportedly beat SkillOpt and CoEvoSkills in all six combinations, even when those baselines received twice the token budget, and reached target performance with 40 to 70% fewer tokens. Paper: https://academy.dair.ai/papers/skilllift-learning-dense-rubrics-from-sparse-oracles-for-efficient-skill-evoluti-2609.15396

#skilllift#agent-skills#research#llm-agents

EvolveTrade rewrites a trading agent's system prompt after each interval

The EvolveTrade paper treats a trading agent's system prompt as its policy and has a separate Policy Agent rewrite it after each interval from decision traces and realized returns, while the backbone model stays frozen. Across several market regimes and two backbone models, it reports higher Sharpe ratios and cumulative returns than fixed-prompt baselines in most settings.

Only 35.8% of the cleanest public terminal-agent RL collection passed a Salesforce AI Research audit

Salesforce AI Research put that collection at 35.8% clean and two others at 10.1% and 3.3%, after finding reward errors in both directions. The authors say RIVER, which filters defective environments, increases RL gains by 106% on Terminal-Bench-Lite and 30% on Terminal-Bench v2.1 while using fewer than 30% of TMax's environments.