Introducing SkillOpt — an optimizer for agent skills.

Instead of finetuning model weights, we treat a natural-language skill as a trainable external parameter.

Think of it as deep learning for the frontier-model + agent era: learning rate, LR schedule, mini-batch, batch size, epoch, momentum — all in text-space optimization. SkillOpt enables stable, controllable skill updates through bounded edits, allowing the optimizer to summarize “gradient directions” from agent experience and continuously improve procedural capability. We evaluate SkillOpt across 6 benchmarks and 7 models, under both direct model calls and real agent execution loops with Codex + Claude Code. SkillOpt achieves best or tied-best results in 52/52 settings.

Train the skill, not the model.

https://aka.ms/skillopt https://huggingface.co/papers/2605.23904…

Executive Strategy for Self-Evolving Agent Skills

Source: https://microsoft.github.io/SkillOpt/ Project Video

SkillOpt in motion.

A short visual overview of how SkillOpt treats natural-language skills as trainable artifacts: roll out, reflect, edit, validate, and export.

Promotional video for the SkillOpt project page. The static paper teaser is shown below for high-resolution inspection.

Paper Teaser

The core loop at a glance.

The teaser summarizes the SkillOpt training loop: rollout evidence, optimizer-side reflection, bounded skill edits, validation gating, and the exported reusable skill.

Figure from the SkillOpt paper. On small screens, the figure area scrolls horizontally to preserve the original details.

A skill is external state for an agent.

Instead of fine-tuning a model or hand-maintaining prompts, SkillOpt runs the frozen agent on scored batches, asks a separate optimizer model to propose structured edits, and accepts a candidate only when validation performance improves.

Frozen target modelOptimizer modelAdd / delete / replace editsHeld-out gate

Rollout

The target model executes tasks with the current skill and records scored trajectories.

Reflect

The optimizer analyzes success and failure minibatches to find reusable procedures.

Edit

Candidate add, delete, and replace operations are merged and ranked under a budget.

Gate

The candidate skill is kept only if it improves held-out selection performance.

Evidence

Rollout batches capture messages, tool calls, verifier feedback, task metadata, and final scores.

Minibatches

Failures and successes are reflected separately so edits correct recurring errors while preserving working behavior.

Bounded Edits

An edit budget functions as a textual learning rate, preventing useful rules from being overwritten by broad rewrites.

Memory

Rejected edits, slow update, and optimizer-side meta skill provide longer-horizon feedback without bloating deployment.

SkillOpt pipeline from the paper. The frozen target model executes with the current skill; the optimizer model proposes bounded edits; held-out validation decides whether the candidate becomes the new current skill. Method comparison

SkillOpt clears the strongest baseline on every benchmark.

ComponentSettingSearchQASpreadsheetLiveMathLearning ratelr=4 default87.177.561.3Learning ratewithout lr84.675.757.3Rejected bufferwith buffer87.177.561.3Rejected bufferwithout buffer85.572.958.9Update memorymeta skill + slow update87.177.561.3Update memorywithout both86.355.059.7

What the ablations say

BoundedTextual learning rates prevent destructive rewrites while keeping enough plasticity to learn new procedures.

GatedHeld-out selection turns reflection into propose-and-test optimization rather than unconditional self-editing.

BufferedRejected edits become negative feedback, helping the optimizer avoid repeating harmful directions.

Epoch checkpoint trends from the paper. Selection-best checkpoints are compared with train rollout score and unseen test performance. ALFWorld skill evolution scoresSelection score rises from 68.6 percent to 81.4 percent, while rejected edits are visible as downward candidate points.85%80%75%70%65%basestep 1step 2step 3slowstep 4

Accepted edits become the current skill only after held-out selection improves.Step 3 is rescued by a slow update; Step 4 trains higher but fails selection.

Cross-model+15.2GPT-5.4 LiveMath skill transferred to GPT-5.4-nano on LiveMathBench.

Cross-harness+31.8Codex-trained SpreadsheetBench skill transferred into Claude Code.

Self-optimizer+10.4GPT-5.4-nano used as its own optimizer improved SpreadsheetBench over baseline.

Deployment1 fileThe target model consumes only the final skill, not optimizer memory.

A stronger optimizer model gives the largest gains, but the loop is not merely distillation from a stronger model. Even matched target-as-optimizer settings can discover useful edits when the update is constrained, buffered, and validated.

@misc{yang2026skilloptexecutivestrategyselfevolving,
      title={SkillOpt: Executive Strategy for Self-Evolving Agent Skills}, 
      author={Yifan Yang and Ziyang Gong and Weiquan Huang and Qihao Yang and Ziwei Zhou and Zisu Huang and Yan Li and Xuemei Gao and Qi Dai and Bei Liu and Kai Qiu and Yuqing Yang and Dongdong Chen and Xue Yang and Chong Luo},
      year={2026},
      eprint={2605.23904},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.23904}, 
}