Tencent's SkillAdam lifts agent skill auto-tuning to 28.3% accuracy using a third of the tokens
rohanpaul_ai · x · 2026-10-02
A new Tencent paper, SkillAdam, tackles two failure modes in auto-rewriting agent skill files: rewriters that go in circles, burning tokens as new edits undo fixes that already worked.
The method teaches the rewriter two habits:
- Memory of past fixes: keep a log of what's been fixed to avoid thrashing;
- A brake on big edits: when results are mixed, make smaller, more conservative changes.
Results: 28.3% average accuracy on long shopping and travel-planning tasks vs. 21.7% for SkillOpt, the prior best, using roughly a third as many tokens. Transferable lesson: give any agent-instruction auto-tuning loop a memory of past fixes and a constraint against large rewrites.
Paper: arXiv 2609.08944, "SkillAdam: Stable and Efficient Skill Evolution for Agents"
More from coding & agent
- Connect Claude Code to NotebookLM via MCP to read full docs without burning tokens — nikola_mr64990 · 2026-10-02
- Dev tracks 31 engineering MCPs, half of them CAD-related, in early visual index — burhop · 2026-10-02
- Deel says its AI agent automated back-office work that legacy tools failed at for years — eptwts · 2026-10-02
- JevSpawn: compositional action spaces speed up agentic inference on 8 benchmarks — Haoyang Su · 2026-10-02
- 7 Claude prompts that turn a messy spreadsheet into insights in 15 minutes — Aiden_Tech_Ai · 2026-10-02
- What should 'approving' an AI memory actually mean? A builder's design dilemma — BarcodeCutter · 2026-10-02