AI research wiki
agentic-systems · 2026-05-25 · Tier 2

SkillOpt: a deep-learning-style optimizer for agent skills

SkillOpt: a deep-learning-style optimizer for agent skills

arXiv: 2605.23904 · HF Daily Papers: page · Date: 2026-05-25 Authors: Yifan Yang, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Chong Luo (Microsoft) with co-authors at Shanghai Jiao Tong, Tongji, Fudan Raw: farmer file

TL;DR

SkillOpt is the first systematic, controllable text-space optimizer for agent skills. A separate optimizer model takes scored rollouts and emits bounded add/delete/replace edits on a single skill document; an edit is accepted only when it strictly improves held-out validation. A textual learning-rate budget, a rejected-edit buffer, and slow/meta epoch updates make the training loop reproducible. Zero inference-time overhead at deployment. Across 6 benchmarks, 7 target models, and 3 harnesses (direct chat, Codex, Claude Code), SkillOpt is best or tied on all 52 (model, benchmark, harness) cells. On GPT-5.5 it lifts no-skill accuracy by +23.5 in direct chat, +24.8 inside Codex, +19.1 inside Claude Code. Skills transfer across model scales and execution environments.

SkillOpt training loop (weight-space discipline applied to a text artifact):

  frozen agent ──► rollouts ──► scored by task reward
                                    │
                                    ▼
                          ┌─────────────────────┐
                          │  Optimizer model    │
                          │  add / del / replace│
                          └─────────┬───────────┘
                                    ▼
                          proposed best_skill.md edit
                                    │
                          ┌─────────▼────────────┐
                          │  Validation gate     │ ── reject ──► rejected-edit buffer
                          │  strict improvement? │
                          └─────────┬────────────┘
                                    ▼ accept
                          best_skill.md (slow/meta epoch updates)
                                    │
                                    ▼
                          deployment: zero inference-time overhead

Key claims

  • The frame is: the skill is the external state of a frozen agent, and that external state should be trained with the same discipline as weight-space optimization (learning rate, validation gate, structured updates, rejected-edit buffer).
  • The optimizer is a separate model from the agent itself; it consumes scored rollouts and emits bounded edits on a single best_skill.md.
  • The validation gate is the key control: an edit is accepted only when it strictly improves the held-out validation score. This prevents the optimizer from churning the skill document and making it incoherent.
  • Slow/meta updates at epoch granularity prevent over-fitting to within-epoch sample noise.
  • The benchmark grid is unusually thorough: 52 cells covering 6 benchmarks, 7 target models, and 3 execution harnesses. SkillOpt is best or tied on every cell.
  • Headline numbers on GPT-5.5: +23.5 direct chat, +24.8 inside Codex, +19.1 inside Claude Code.
  • Transfer: a skill optimized on one model and one harness retains value when moved to a different scale or to a different harness, and to a nearby math benchmark without further optimization.
  • Zero deployment overhead because the skill is just a text artifact.

Relation to prior wiki content

SkillOpt is the second paper in May to operationalize the "external state is trainable" frame, joining the Life-Harness thread from 05-23 (the code-as-agent-harness survey that argued the runtime interface, not the model weights, is the right design surface for agent improvement). Life-Harness evolved the harness from trajectories. SkillOpt evolves the skill document from trajectories. Both keep the LLM frozen. Both make the external state the trainable surface. The two papers together establish a clean two-axis taxonomy: harness (the wrapper) and skill (the procedural knowledge), each independently trainable, each independently transferable.

It is also the third paper in May on disciplined text-space optimization, joining GEPA-style prompt optimization and EvoSkill. SkillOpt's contribution over these is the validation gate plus the rejected-edit buffer; both are direct imports from deep-learning optimization, and both are what make the training loop stable. Without them, prior text-space optimizers degenerate as the skill document grows.

The transfer claim is the most consequential. If a skill optimized on Qwen-7B transfers to GPT-5.5 inside Claude Code without retraining, the cost-of-deployment model for agentic systems flips. The skill becomes a portable artifact like a model checkpoint, not a per-model-per-environment fixture. Combined with SDAR (the 05-15 self-distilled agentic RL paper) and the code-as-agent-harness thread, the picture is that agent intelligence is now decomposing into a stable model + portable harness + portable skill stack.

Research angle

The validation gate is the load-bearing mechanism. Without it the optimizer drifts. The paper's gate is binary (strict improvement on validation); a probabilistic gate based on uncertainty would likely be tighter and would suggest a Bayesian-optimization analog for text-space optimization. Worth a follow-up.

A second open question: how does the rejected-edit buffer interact with the meta updates? The paper treats them as orthogonal controls, but a rejected edit at one epoch might be the right edit two epochs later (after other edits have been accepted). A learned policy over the buffer would be the natural next step.

The transfer claim deserves stress testing on real production loads. Microsoft is well-positioned to run that test; the lift on Codex (+24.8) suggests this is being prepared for production deployment in their coding tools.