SkillOpt: Executive Strategy for Self-Evolving Agent Skills

1 Introduction

• SkillOpt uses a separate optimizer model to turn rollout evidence into bounded add/delete/replace edits, accepting updates only when held-out validation performance improves.Candidate edits are aggregated and ranked under a textual learning-rate budget. Deployment uses a roughly 300–2,000-token best_skill.md artifact while the target model and execution harness remain fixed.Source: Figure 1 · p. 2, §1 · pp. 2–3
• SkillOpt treats skill documents as trainable external state, addressing brittle manual or one-shot skills and the expense or unavailability of weight adaptation.Agent skills package procedures, domain heuristics, tool policies, output constraints, and failure modes into portable natural-language artifacts for frozen agents [7] [8].Source: §1 · pp. 1–2
• +23.5, +24.8, and +19.1 points over no skill are GPT–5.5’s gains in direct chat, Codex, and Claude Code, respectively.Direct chat gains average +5.4 points over the strongest per-cell baseline among human-written, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills. Against EvoSkill, gains are +14.0 points in Codex and +3.2 points in Claude Code.Source: §1 · p. 3
• +59.7 points is the gain from transferring a Codex-trained spreadsheet skill to Claude Code, demonstrating reuse beyond its training harness.A SpreadsheetBench skill trained on GPT–5.4 improves every tested smaller GPT variant, while an OlympiadBench skill produces positive gains on Omni-MATH [14]. These artifacts can be audited as text and reused without model-weight changes.Source: §1 · p. 3
• Bounded textual learning outperforms uncontrolled rewriting; validation gating, rejected-edit feedback, and slow/meta updates respectively prevent harmful accumulation, learn from failures, and improve refinement.Rollout/reflection batch sizes control evidence noise, learning-rate budgets and schedules constrain edit size, and slow/meta updates carry stable editing directions across epochs. Case studies produce compact, procedural rather than instance-specific skills after only 1–4 accepted edits.Source: §1 · pp. 2–3
• 52 of 52 evaluated cells show SkillOpt best or tied-best across six benchmarks, seven target models, and three execution harnesses.Comparisons include no-skill, human-written, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill baselines.Source: §1 · p. 3

2 Related Work

• GEPA evolves prompts using trajectory feedback [13], while ABSTRAL and EvoTest extend this approach to multi-agent design documents and test-time agentic system evolution [15] [16].GEPA outperforms reinforcement learning on several language-agent tasks [13]. These approaches primarily optimize prompts, system designs, or full configurations rather than reusable domain adaptation.Source: §2 · p. 3
• SkillsBench and the SoK frame skills as reusable procedural knowledge [7] [8]; prior systems construct [9] [12] [17] [18] [19] and refine skills [10] [11] [20] [21] [22] [23] [24] [25] [26].Construction draws on lifelong experience, trajectory lessons, knowledge bases, and domain resources [9] [12] [17] [18] [19]. Refinement includes failure analysis, creation-evaluation-revision loops, coevolving generators and verifiers, collective updates, and reinforcement learning [10] [11] [20] [21] [22] [23] [24] [25] [26].Source: §2 · p. 3, EPOCH-WISE SLOW/META UPDATE · p. 4

EPOCH-WISE SLOW/META UPDATE

• SkillOpt trains one compact, persistent domain skill with deep-learning-style controls, producing a portable best_skill.md without changing model weights.Controls include trajectory batches, reflection minibatches, textual learning rates, validation gates, rejected-edit buffers, and slow/meta updates. The skill document can be trained, validated, exported, and reused.Source: §2 · p. 3, EPOCH-WISE SLOW/META UPDATE · p. 4

3.1 Problem Setup

• SkillOpt optimizes an injected natural-language skill while freezing the target model, using training experience, selection-gated updates, and a test split reserved for final reporting.The skill is prepended to system or developer instructions in direct chat and becomes persistent procedural memory in tool-use harnesses. Optimizer state tracks current and best skills, cached hashes, rejected steps, and optional slow/meta state; only the best accepted skill is exported as best_skill.md.Source: §3.1 · pp. 4–5

3.2 Forward Pass: Rollout Evidence

• Rollout batches supply scored execution evidence, with batch size controlling the trade-off between update speed and exposure to recurring patterns.Evidence includes messages, tool interactions, outputs, answers, verifier feedback, and benchmark-specific context. Optional accumulation reflects on several batches separately and merges them into one update, decoupling execution throughput from update frequency.Source: §3.2 · p. 5

3.3 Backward Pass: Minibatch Reflection

• Reflection separates successful and failed trajectories into minibatches, then hierarchically merges proposals with priority on failure corrections.Failures propose corrective rules, while successes preserve working behaviors; outputs are structured add/delete/replace edits or rewrite suggestions. Merging filters duplicate, contradictory, and example-specific proposals before selecting a bounded update.Source: §3.3 · p. 5

3.4 Bounded Text Updates

• An edit budget Lt bounds each update by retaining the highest-ranked proposals, with the default cosine schedule decreasing edit counts over training.Constant, linear, and autonomous schedules are also supported. Selected proposals become localized patches or condition a full rewrite, but step-level updates cannot overwrite the protected slow-update field.Source: §3.4 · p. 5

3.5 Validation Gate and Rejected-Edit Buffer

• Candidate skills are accepted only when they improve the current selection score; rejected edits inform subsequent reflection within the same epoch.Evaluation uses the same frozen target model and harness. The rejected-step buffer records failure patterns, attempted edits, and resulting score drops, providing training feedback without inference-time cost.Source: §3.5 · p. 6

3.6 Epoch-Wise Slow/Meta Update

• Epoch-wise updates compare previous and current skills on identical training items to produce validation-gated longitudinal guidance.Comparisons distinguish improvements, regressions, persistent failures, and stable successes, informing a protected slow-update field. Separate optimizer-only meta guidance summarizes helpful and rejected edit patterns for future reflection, merging, and ranking; it is not deployed.Source: §3.6 · p. 6

3.7 Harness-Agnostic Deployment

• Lightweight adapters make SkillOpt harness-agnostic by constructing batches, injecting skills, executing native harnesses, and returning scored trajectories.The interface supports direct QA, spreadsheet execution, document reasoning, multimodal QA, embodied environments, and Codex-style or Claude Code-style loops. Offline-trained skill artifacts can be deployed or tested across target models, harnesses, and nearby benchmarks without changing model weights.Source: §3.7 · p. 6

4 Experiments

• Six benchmarks evaluate frozen agents against seven baselines using matched target models, held-out test splits, and scorers across factual, procedural, and embodied tasks.Baselines comprise no skill, human skill, one-shot LLM skill, Trace2Skill [9], TextGrad [40], GEPA [13], and EvoSkill [10]. Seed-42 splits reserve selection data for edit acceptance and disjoint tests for reporting; direct chat prepends skills, while Codex and Claude Code share the skill-file contract and expose execution traces for reflection.Source: §4 · pp. 6, 8, 9

4.1 Main Results

• 52/52 evaluated model–benchmark–harness cells show SkillOpt matching or beating the strongest competing baseline across seven target models and three harnesses.GPT–5.5 direct-chat accuracy averages 82.3 versus 58.8 without skills (+23.5 points) and 76.9 for the per-cell oracle baseline (+5.4). Average improvement across the seven direct-chat models is approximately +17.6 points.Source: Table 1 · p. 7, §4.1 · pp. 9–11
• +24.8 points under Codex and +19.1 under Claude Code are GPT–5.5’s average gains over no skill, winning all five evaluated benchmarks in each harness.Average margins over EvoSkill are +14.0 points under Codex and +3.2 under Claude Code; on Codex SpreadsheetBench, SkillOpt adds +17.5 points over EvoSkill (67.5→85.0). ALFWorld is excluded from both harness evaluations because their adapters do not support its persistent embodied interaction.Source: Table 1 · p. 7, §4.1 · pp. 10–11

4.3 Analysis and Transfer

• Optimized skills transfer positively across model scales, Codex/Claude Code harnesses, and OlympiadBench [41] to Omni-MATH [14], without further optimization.SpreadsheetBench transfer gains are +59.7 points from Codex to Claude Code and +43.6 in reverse; transferred LiveMath skills score 28.8 on GPT–5.4-nano versus 27.2 in-domain. OlympiadBench→Omni-MATH gains are +3.7, +1.8, and +1.3 points on GPT–5.4, GPT–5.4-mini, and GPT–5.4-nano, respectively.Source: Table 4 · p. 9, §4.3 · pp. 12–13

4.2 Ablations

• 47.5→78.0 on SpreadsheetBench and 59.1→70.5 on LiveMathematicianBench accompany increasing training evidence, while batch-size choices are comparatively robust.SearchQA saturates around 84–86 after 20% of training data; reflection minibatches from 1–32 and rollout batches from 8–a full epoch produce relatively stable scores. Small or moderate bounded edit budgets remain competitive across constant, cosine, and linear schedules.Source: Table 2 · p. 8, §4.2 · p. 11
• −22.5 points on SpreadsheetBench (77.5→55.0) follow removing both slow update and meta skill; removing rejected-edit memory also reduces performance.The strict validation gate rejects ties, step-level edits cannot overwrite the slow update’s protected region, and the optimizer-side meta skill remains teacher-only. Validation checkpoints track held-out test performance across epochs, supporting the gate’s ability to select generalizing skills.Source: Table 3 · p. 8, §4.2 · pp. 11–12, Figure 3 · p. 12

4.3 Analysis and Transfer

• 56–74% of frontier-optimizer gains are recovered by target-matched optimizers, while the stronger optimizer improves every tested cell without adding deployment-time calls.For GPT–5.4-nano, frontier versus target-matched optimizer gains are +19.0 versus +11.9 on SpreadsheetBench and +19.0 versus +14.1 on SearchQA. Both regimes retain identical rollout batches, validation gating, bounded edit budgets, rejected-edit memory, and slow/meta updates.Source: §4.3 · p. 13, Table 5 · p. 14

4.4 Learned Skills: Compactness, Cost, and Examples

• 379–1,995 tokens, with a median roughly 920, comprise final skills produced through only 1–4 accepted edits across six benchmarks.A single accepted edit produces LiveMathematicianBench’s +29.3-point gain and OfficeQA’s +39.0-point gain over no skill. Example rules prescribe choosing short canonical answers supported by distinctive evidence and inspecting workbook formulas before writing evaluated static values across requested ranges.Source: §4.4 · pp. 14–15, Table 6 · p. 14
• 0.6–3.6M training tokens per absolute test-set point suffice for procedural benchmarks, versus 37.9M for SearchQA and 46.4M for DocVQA.OfficeQA gains +39.0 points at 1.1M tokens per point. Longer trajectories or richer multimodal context make SearchQA and DocVQA roughly an order of magnitude more expensive per point.Source: §4.4 · p. 15
• Learned rules are procedural rather than instance-specific, automatically generated, and validated edit-by-edit on held-out data.Examples enforce answer formats in OfficeQA and LiveMathematicianBench, visual-region evidence binding in DocVQA, workbook-structure-first reasoning in SpreadsheetBench, search-frontier discipline in ALFWorld, and canonical-entity choice in SearchQA. The authors interpret these generalizable rules as evidence of transferable procedural knowledge rather than merely training-split overfitting.Source: §4.4 · p. 15
• < 2,000 tokens and 1–4 accepted edits characterize the compact, interpretable deployed skills.Training costs are paid once; exported skills add no optimizer calls or weight updates, only a compact text artifact.Source: §4.4 · p. 15

4.5 Qualitative Skill Evolution

• 49.3 to 74.6: ALFWorld held-out test performance improves in a representative run as edits add object identity, search memory, progress locks, and loop breakers.The run uses GPT–5.4-nano as student and GPT–5.5 as teacher. Exact object-name matching, visited-location and destination memory, pick-two progress locks, and immediate admissible completion actions turn the generic search-transform-place plan into a finite-state execution policy.Source: §4.5 · p. 16
• 40.4 to 78.9: SpreadsheetBench held-out test performance improves in a representative run through workbook inspection, data normalization, formatting preservation, and output validation.GPT–5.5 serves as both frozen student and optimizer; edits require inspecting actual workbooks and locating headers and ranges across sheets. Rules require evaluated static values when graders read cell values, complete target ranges, Python-only helper computations, and reopening saved workbooks to check boundary rows and blanks.Source: §4.5 · p. 16

5 Conclusion

• 52 of 52 evaluated cells show SkillOpt best or tied-best across six benchmarks, seven target models, and three execution modes.GPT–5.5 gains over no skill average +23.5 points in direct chat, +24.8 under Codex, and +19.1 under Claude Code. SkillOpt beats the strongest per-cell baseline among human, LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills by +5.4 points on average.Source: §5 · p. 16
• SkillOpt makes an external skill document trainable while freezing the target agent and separating task execution from optimizer-driven skill editing.Bounded edit budgets, minibatch reflection, held-out validation gates, rejected-edit buffers, and epoch-wise slow/meta updates make improvement a controlled learning process rather than ad hoc prompt revision.Source: §5 · p. 16
• 1–4 accepted edits assemble compact (< 2,000 token), interpretable skill artifacts that case studies show transfer across model scales, harnesses, and nearby benchmarks.The authors suggest these reusable skills can provide practical domain adaptation without modifying model weights.Source: §5 · p. 16

Outlook

• SkillOpt currently optimizes one skill for one domain; proposed extensions include cross-domain libraries, reusable optimizer meta skills, alternative validation gates, and skill self-distillation.Reward-free or preference-driven gates could support open-ended tasks, while self-distillation into the target model could lead toward weight-level adaptation. Future work could apply learning rates, schedules, regularization, curricula, and validation to skill optimization.Source: §5 · p. 17

• SkillOpt keeps the task-execution model fixed and trains only a text skill document.Source: A Additional Method Details and… · p. 20
• A separate optimizer reads rollout evidence, proposes patch-style edits, merges and ranks them, and submits each candidate skill to a held-out selection gate.Source: A Additional Method Details and… · p. 20
• The task-execution model receives only the current skill and benchmark task, without access to the optimizer prompts.Source: A Additional Method Details and… · p. 20

B Limitations

• SkillOpt requires scored trajectories and held-out selection, making it most directly applicable to tasks with reliable feedback.Subjective, multidimensional, or costly-to-judge outcomes may require stronger human or model-based evaluation for the validation gate.Source: B Limitations · p. 20
• Skill training requires additional rollout computation and optimizer-model calls, which may be less attractive for one-off tasks.Deployment uses only a compact best_skill.md artifact; training costs are amortized when the same skill is reused.Source: B Limitations · p. 20
• A single portable skill simplifies deployment but may be insufficient for heterogeneous domains requiring many disjoint procedures.SkillOpt deliberately avoids growing a large skill library or changing model weights.Source: B Limitations · p. 20
• Optimized skills may encode training-distribution heuristics, requiring careful held-out evaluation before transfer to substantially different models, harnesses, or task settings.Source: B Limitations · p. 20

Benchmarks and metrics

• Native benchmark evaluators measure hard success or exact-match accuracy on held-out tests; a separate selection split is used only to select candidate skills.SearchQA covers extractive QA; SpreadsheetBench, spreadsheet code and tool use; OfficeQA and DocVQA, document reasoning; SealQA, noisy retrieval; LiveMathematicianBench, mathematical multiple-choice reasoning; and ALFWorld, sequential decisions. Dataset-backed benchmarks use deterministic train/selection/test splits, defaulting to 2:1:7 unless otherwise specified.Source: C Experimental Protocol Details · p. 20

Baselines

• Baselines include no skill, manually written skills, one-shot generated skills, Trace2Skill, TextGrad, GEPA, and EvoSkill, with applicability constrained by execution setting and aligned evaluation.Trace2Skill mines training trajectories without SkillOpt’s iterative validation gate; TextGrad and GEPA are reflective prompt optimizers for direct chat. EvoSkill appears in harness-backed comparisons where a matched completed run exists, and entries not measured under the final aligned protocol are marked –.Source: C Experimental Protocol Details · pp. 20–21

Optimization protocol

• Candidate skill edits are accepted only when they improve the current selection score, and the best accepted skill is exported as best_skill.md.The optimizer analyzes successes and failures separately, proposes patch-style edits, merges duplicate or contradictory proposals, and ranks edits under the learning-rate cap. The student model, backend, harness, and benchmark evaluator remain fixed.Source: C Experimental Protocol Details · p. 21
• Four epochs, rollout batch size 40, reflection minibatch size 8, and textual learning rate 4 define the default optimization configuration.Defaults also include cosine learning-rate decay with minimum rate 2, held-out validation gating, slow update enabled with 20 sampled examples, and optimizer-side meta skill enabled.Source: C Experimental Protocol Details · p. 21

Ablation protocol

• One-factor ablations change a single scalar or component while holding the remaining optimizer configuration fixed.Train-size ablations retain a 2:1:7 train/selection/test split and vary training exposure; the 100% row uses the full training partition and is directly comparable to smaller subsets. Component ablations alter the edit budget, rejected-edit buffer, and epoch-wise slow/meta update individually.Source: C Experimental Protocol Details · p. 21

C.1 Optimization Procedure

• Optimization tracks the current skill, validation-gated best skill, selection-score cache, rejected edits and failure patterns, plus optimizer-side meta guidance for future edits.The meta skill guides future edit generation rather than serving as training-model-facing instructions.Source: §C.1 · p. 21, §C.2.8 · p. 26

C.2 Optimizer Prompt Contracts

• Optimizer prompt contracts require JSON outputs so edits can be parsed, filtered, applied, and validated without manual intervention.The reproduced operational contracts normalize terminology to the paper’s optimizer/training-model framing.Source: §C.2 · p. 21

C.2.3 Failure merge: merge_failure.md

• Failure merging deduplicates corrective edits, resolves conflicts by justification, prioritizes recurring patterns, and preserves nonredundant insights without overlapping edit targets.Task-specific singleton edits may be discarded; each retained edit receives an estimated source-patch support count. Content between the slow-update markers is protected from edits.Source: §C.2.3 · pp. 23–24

C.2.4 Success merge: merge_success.md

• Success merging conservatively adds only generalizable patterns absent from the skill, prioritizing patterns recurring across successful trajectories.Similar patterns are deduplicated, each edit receives an estimated source-patch support count, and the protected slow-update section cannot be modified.Source: §C.2.4 · p. 24

C.2.5 Final merge: merge_final.md

• Final merging prioritizes failure corrections over success reinforcement, preserves complementary success insights, and carries edit support counts and provenance forward.Failure edits survive unless directly conflicting with well-supported success patterns; duplicates retain the failure version, and edits surviving earlier merge rounds receive priority. Allowed operations are append, insert_after, replace, and delete, but cannot target the protected slow-update section.Source: §C.2.5 · p. 25

C.2.6 Ranking and selection: ranking.md

• Edit selection follows a supplied budget, ranking proposals by systematic impact, complementarity, generality, and actionability, in that order.Selected edits are returned as zero-based indices in priority order, alongside a brief ranking justification.Source: §C.2.6 · p. 25

C.2.7 Slow update: slow_update.md

• Slow updates use same-task comparisons across consecutive epochs to rewrite protected guidance that step-level optimization cannot modify.The comparison uses the same 20 training tasks, categorized as regressions, persistent failures, improvements, or stable successes. Updated guidance retains effective prior instructions, revises ineffective ones, complements the main skill, and prioritizes preventing regressions, fixing persistent failures, then reinforcing successes.Source: §C.2.7 · p. 26

C.2.8 Optimizer memory: meta_skill.md

• Optimizer memory provides compact guidance for future edit generation, addressing the optimizer rather than supplying task instructions to the training model.Adjacent-epoch evidence identifies helpful edit types, suitable abstraction levels, failure-repair priorities, and risks from vague, redundant, brittle, or harmful edits. Memory favors durable principles and revises or removes ineffective earlier guidance.Source: §C.2.8 · pp. 26–27

C.2.1 Failure analysis: analyst_error.md

• The analyst reads every minibatch trajectory, classifies systematic failures, and proposes generalizable edits that fill skill gaps without duplicating existing content.At most L edits address the highest-impact patterns, with fewer permitted. Content between SLOW_UPDATE_START and SLOW_UPDATE_END markers is protected from modification or deletion.Source: §C.2.1 · pp. 21–23
• SkillOpt separately analyzes failure and success minibatches, merges proposals with failure priority, ranks edits within the scheduled budget, and applies them to a candidate skill.Source: §C.2.1 · p. 22
• Candidate skills replace the current skill only after strict score improvement; rejected edits and observed failure patterns enter a buffer.Epoch-end comparisons use the same sampled tasks under previous and current skills to generate protected longitudinal guidance, validated through Dsel. The optimizer also updates meta guidance for future edit generation and selection.Source: §C.2.1 · p. 22
• The analyst returns JSON containing batch size, failure types, counts and descriptions, edit rationale, and structured edit operations.Supported operations are append, insert_after, replace, and delete, with exact targets where required. The edits list may be empty when no patch is warranted.Source: §C.2.1 · pp. 22–23

• The analyst examines multiple successful trajectories from one minibatch alongside the current skill to identify shared, generalizable behaviors worth encoding.Source: §C.2.2 · p. 23
• Proposed patches must address uncovered patterns recurring across multiple trajectories, remain concise and task-general, and preferably reinforce existing sections rather than add top-level sections.Source: §C.2.2 · p. 23
• At most L edits may be proposed, prioritizing broadly applicable patterns and using fewer edits when warranted.Source: §C.2.2 · p. 23
• The JSON response records batch size, success patterns, patch reasoning, and edits using append, insert_after, replace, or delete operations.Edits may be empty when the skill already covers all observed patterns. Edits must not target, modify, or delete content between the protected SLOW_UPDATE_START and SLOW_UPDATE_END markers, which a separate slow-update process manages.Source: §C.2.2 · p. 23

• Patch-mode updates permit only append, insert_after, replace, and delete operations.Merged edits record support counts and source types, allowing ranking to favor edits surviving independent analyses and hierarchical merges.Source: §C.3 · p. 27
• The edit budget Lt acts as a textual learning rate, limiting applied edits per step to preserve continuity between adjacent skills.Source: §C.3 · p. 27
• Only epoch-boundary slow updates may rewrite the protected section, and rewritten skills must pass the held-out selection gate before becoming current.SLOW_UPDATE_START and SLOW_UPDATE_END delimit the section, which all step-level prompts are prohibited from changing.Source: §C.3 · p. 27
• Rejected edits and their failure patterns remain in the step buffer so later optimizer calls can avoid repeating harmful changes.Source: §C.3 · p. 27

C.4 Design Principles

• The task-execution model remains fixed; only the text skill changes.Source: §C.4 · p. 27
• Every candidate skill is evaluated on a selection split before acceptance, preventing unvalidated reflection from accumulating.Source: §C.4 · p. 27
• Minibatch analyses are merged hierarchically so final edits reflect recurring evidence rather than single examples.Source: §C.4 · p. 27
• The edit budget acts as a learning-rate analogue, allowing larger early changes and smaller late refinements.Source: §C.4 · p. 27
• The deployed skill stays lightweight and inspectable, while the optimizer-side meta skill remains separate from the skill shown to the task-execution model.Source: §C.4 · p. 27