COBRA-Skills: contextual bandits cut agent skill optimization cost by ~55-58%
GardenDelicious1476 · reddit · 2026-09-13
COBRA-Skills, an open-source research project, frames agent skill optimization as a sequential budget-allocation problem — deciding which skill is worth evaluating next. It keeps a population of candidate skills, estimates utility with a neural predictor, uses LinearUCB to balance exploration/exploitation, evaluates only selected skills on the target agent, and reuses execution feedback for scheduled regeneration, mutation, and crossover. Across 6 agent benchmarks (search QA, spreadsheets, document understanding, math, social reasoning, embodied) and 3 target models: best average performance, 55-58% lower cost than SkillOpt, only 50 optimization examples per benchmark, works under Codex and Claude Code harnesses. Key finding: most savings come from avoiding repeated LLM-based trajectory analysis and skill rewriting, not fewer target-agent executions. Paper: arXiv:2609.11682; code and HF page open.
Related event: COBRA-Skills Cuts Agent Skill Optimization Costs by Over 50%(2 posts)→
More from coding & agent
- Gym-Anything Adds Audit Agents That Try to Break Simulations With Legal Robot Actions — wellecks · 2026-10-30
- CMU's Gym-Anything Auto-Builds 200+ Agent Environments from Any Software, Selected by GDP — wellecks · 2026-10-30
- durable-object-migrator: zero-downtime exports for Cloudflare Durable Objects — samgoodwin89 · 2026-10-30
- This 5-question interview prompt makes Grok design your first two custom bots in minutes — nateliason · 2026-10-30
- Cursor and Harvey swap general models for task-specific brains, but most niches can't afford to — MartinGTobias · 2026-10-30
- Scenario open-sources 3D OS: 80 agent skills taking sketches to game-ready assets — Scobleizer · 2026-10-30