Source-linked AI summary
SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, Xiangxiang Chu
TL;DR
SkillClaw addresses static skills and isolated improvements by aggregating multi-user trajectories into shared evidence and using an agentic evolver to update a common repository. Validated updates are synchronized across users, and experiments report substantial task improvements, while the study remains a small-scale test with limited interaction signals and depth.
Problem
Existing systems do not reliably convert heterogeneous cross-user experiences into accumulating skill improvements, leaving skills static and repeated failures isolated.
Method
SkillClaw structures and aggregates interaction trajectories, uses an open-ended agentic evolver to refine or create skills, validates candidate updates, and synchronizes accepted skills.
Results
SkillClaw yields substantial improvements across WildClawBench tasks for qwen3-max in a simulated multi-user deployment, including Search & Retrieval rising from 22.73% to 34.55% and Creative Synthesis from 11.57% to 21.80%.
Takeaways & Limitations
SkillClaw turns ordinary usage into shared, continuously evolving skills whose validated improvements can propagate across users without additional user effort.
Takeaways & Limitations
The study is a small-scale test with limited user queries, feedback signals, and interaction depth.
Abstract
from arXiv · showhide
Large language model (LLM) agents such as OpenClaw rely on reusable skills to perform complex tasks, yet these skills remain largely static after deployment. As a result, similar workflows, tool usage patterns, and failure modes are repeatedly rediscovered across users, preventing the system from improving with experience. While interactions from different users provide complementary signals about when a skill works or fails, existing systems lack a mechanism to convert such heterogeneous experiences into reliable skill updates. To address these issues, we present SkillClaw, a framework for collective skill evolution in multi-user agent ecosystems, which treats cross-user and over-time interactions as the primary signal for improving skills. SkillClaw continuously aggregates trajectories generated during use and processes them with an autonomous evolver, which identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide while requiring no additional effort from users. By integrating multi-user experience into ongoing skill updates, SkillClaw enables cross-user knowledge transfer and cumulative capability improvement, and experiments on WildClawBench show that limited interaction and feedback, it significantly improves the performance of Qwen3-Max in real-world agent scenarios.
1 Introduction
SkillClaw addresses the limitation that agent skills remain static and improvements stay isolated by aggregating cross-user interaction evidence and evolving a shared skill repository automatically.
- Motivation: Existing agent adaptations store instance-specific memories or static skill instructions, limiting accumulation of improved behavior across users and time.Local refinements remain isolated, producing fragmented rather than collective improvement.
- Approach: SkillClaw aggregates trajectories from everyday multi-user usage and processes them centrally to drive shared skill updates.The framework uses interaction sessions as evidence for evolution across users and over time.
- Design properties: The framework supports collective, fully automatic, and agentic evolution through shared knowledge, background synchronization, and open-ended reasoning.Updates can refine existing skills or create new capabilities without manual curation or explicit user intervention.
- Evaluation: SkillClaw is designed for multiple Claw-style systems and improves Qwen3-Max performance across WildClawBench tasks in simulated multi-user deployment.The evaluation uses WildClawBench and qwen3-max as the backbone model.
2 Method
SkillClaw converts isolated interaction sessions into shared evidence, uses an agentic evolver to update skills conservatively, and synchronizes validated improvements across agents.
- From isolated sessions to shared evidence: SkillClaw aggregates diverse user interactions because shared evidence helps distinguish generalizable improvements from idiosyncratic fixes.Different users expose complementary conditions under which a skill works or breaks.
- Session structuring: Structured trajectories preserve prompts, actions, tool calls, intermediate feedback, errors, and final responses so procedural failures can be diagnosed.Metadata records referenced skills, tool errors, and coarse quality estimates without imposing rigid labels.
- Evidence organization: Sessions are grouped by referenced skills, enabling comparisons across users, tasks, and environments and identifying recurring success and failure patterns.Sessions without skill usage are placed in G(∅) for discovery of missing reusable procedures.
- Agentic skill evolution: The agentic evolver analyzes grouped successful and failed executions, then chooses to refine, create, or skip an update based on evidence.Successful sessions preserve effective invariants while failed sessions identify behaviors requiring correction.
- Synchronization and evolution loop: Candidate updates undergo validation before merging, and the updated repository is synchronized to agents to generate evidence for the next evolution round.The loop connects session recording, centralized processing, repository updates, and synchronization.
- System properties: Collective evolution, full automation, and agentic adaptability allow shared knowledge to benefit users and address previously unseen failure modes.Users provide only normal agent usage while evolution occurs in the background.
3 Experiments
Across four capability categories, SkillClaw progressively consolidates validated skill updates, with improvements following category-specific bottlenecks rather than a uniform trajectory. Controlled validation and task examples further show that evolution can improve execution reliability and correctness.
- Main Results: SkillClaw consolidates nightly validated updates into a stable skill pool over six days rather than producing daily performance fluctuations.The validator retains locally effective updates for daytime deployment.
- Main Results: Social Interaction rises from 54.01% to 60.34% on Day 2 and remains stable thereafter.The early gain is associated with workflow organization, cross-source integration, and high-level summarization.
- Main Results: Search & Retrieval increases from 22.73% to 30.00%, then to 34.55%, through staged improvements from input reliability to retrieval planning.The progression first addresses file accessibility and validation before higher-level constraint-aware planning.
- Main Results: Creative Synthesis jumps from 11.57% to 21.80% on Day 2 and then plateaus as evolution first resolves environment and multimodal pipeline setup.More complex multimodal skills emerge and pass validation but do not surpass the early best pool within six days.
- Main Results: Safety & Alignment improves from 24.00% to 32.00%, with updates emphasizing reliable and recoverable execution under real-world constraints.Examples include Git fallback, directory cloning protocols, and safe execution in non-interactive settings.
- Analysis: Category-specific bottlenecks shape evolution: executability, input reliability and planning, multimodal pipeline organization, and recoverable execution.These trajectories show that skill evolution is structured rather than simple rule accumulation.
- Controlled Validation: A controlled validation on three failure-focused queries yields a consistent average gain of +42.1% after one evolution round.The queries isolate basic extraction, deadline parsing, and report-saving failures.
- Case Studies: Skill evolution restructures task pipelines as well as correcting errors, including preview-based Slack analysis and more precise document-affiliation reasoning.The examples contrast naive or heuristic workflows with structured evolved procedures.
4 Related Work
Prior work advances agent adaptation through reflection, reusable skills, and experience accumulation, but SkillClaw performs evolution at the group level by aggregating sessions from distributed agents.
- Agent Self-Evolution: Agent self-evolution has progressed from local reflection on individual trajectories toward broader experience accumulation and autonomous improvement.Prior studies include self-correction, reusable lessons, and contextual replay, but mainly improve agents from their own histories or within single optimization loops.
- Collective Evolution: SkillClaw differs by performing evolution at the group level, aggregating sessions from distributed local agents.This positions cross-user session aggregation as the framework’s distinctive evolution setting.
- Skill-Based Adaptation: Skill-based approaches encode standardized procedures and study skill-library accumulation, optimization, discovery, refinement, transfer, and routing.Related work also explores web skill induction, automated multi-agent discovery, recursive skill-augmented learning, evolving memory skills, and lifelong skill self-evolution.
5 Conclusion
SkillClaw turns ordinary interaction trajectories into shared evidence for agentic skill updates through refinement and creation. This enables knowledge discovered during usage to accumulate and propagate across users over time.
- Contribution: SkillClaw presents a framework for collective skill evolution in multi-user agent ecosystems.The framework targets shared evolution rather than isolated skill behavior.
- Contribution: Interaction trajectories become shared evidence that an agentic evolver uses to refine existing skills and create new ones.The resulting updates let knowledge discovered during usage accumulate and propagate across users over time.
- Implications: The framework shifts skills from static libraries toward dynamic, interaction-driven ecosystems that capture recurring patterns, correct failures, and adapt to diverse execution environments.This continuous evolution loop connects interaction-level improvements with system-level capability growth.
Summarize Session Prompt
The session prompt directs the evolver to analyze agent trajectories and make evidence-based skill-library decisions. It emphasizes targeted edits, preservation of environment-specific guidance, and structured outputs for improving, optimizing, creating, or skipping skills.
- Session analysis: The evolver summarizes each session’s goal, trajectory, skill effectiveness, turning points, tool usage, and outcomes before deciding on changes.The prompt requires a trajectory-aware analysis of what the agent tried, why failures occurred, and which skills helped or hurt.
- Evolution actions: The system supports four actions: improve_skill, optimize_description, create_skill, and skip, each with a prescribed rationale and skill payload or decision.New skills should address distinct, recurring, teachable capabilities; description optimization changes only triggering language, while skip is valid when evidence is insufficient.
- Editing principles: Skill changes must respond to recurring evidence while preserving the current skill’s purpose, structure, effective guidance, and environment-specific details.The prompt favors targeted edits over rewrites and cautions against casually changing APIs, ports, endpoints, paths, payloads, filenames, or unrelated examples.
- Failure diagnosis: Failures must be classified as skill, agent, or environment problems so agent-runtime issues and unstable infrastructure do not unnecessarily bloat skills.Correct skill guidance should not be replaced merely because an agent failed to use it, while repeated environment instability may warrant a brief note.
- Workspace procedure: The evolver operates within a workspace containing session inputs, skill-library outputs, persistent history, and a manifest, while following strict file-access and update constraints.It must read all sessions, inspect current skills and relevant history, write skill files autonomously, and leave read-only inputs and server-managed manifests unchanged.