SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
| Authors | Ziyu Ma et al. |
| Year | 2026 |
| HF Upvotes | 267 |
| arXiv | 2604.08377 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
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.
Engineering Breakdown
Plain English
SkillClaw addresses a critical problem in multi-user LLM agent systems: skills (reusable components that tools and workflows) remain static after deployment, forcing each user to rediscover the same failure modes and optimal patterns independently. The paper proposes a framework that treats interactions across all users and time periods as training signals to collectively evolve skills, aggregating heterogeneous trajectories to identify when skills succeed or fail across different contexts. Rather than maintaining a frozen skill library, SkillClaw continuously refines skills based on real-world usage patterns, enabling the system to improve from collective experience rather than individual trial-and-error. This approach fundamentally shifts agent systems from static libraries toward living, adaptive components that compound improvements across the entire user ecosystem.
Core Technical Contribution
The core technical novelty is a skill aggregation and refinement mechanism that transforms distributed, heterogeneous user interactions into reliable skill updates without centralized retraining. SkillClaw introduces a framework for detecting when skills fail across different user contexts, clustering similar failure modes, and propagating fixes back to the shared skill library—essentially treating the multi-user agent system as a distributed learning environment. The key architectural insight is treating skill evolution as a continuous, consensus-building process where cross-user trajectories vote on whether a skill should be modified, rather than requiring explicit human labeling or centralized model updates. This differs from prior work (like OpenClaw) which assumes skills are frozen once deployed, and from standard RLHF which requires synchronized, centralized training—SkillClaw enables asynchronous, federated skill improvement across autonomous agents.
How It Works
SkillClaw operates on a three-stage pipeline: (1) Trajectory Collection—each user interaction with an agent generates a trajectory (sequence of skill invocations, tool calls, outcomes) which is tagged with success/failure signals; (2) Skill Aggregation—the system groups trajectories across users by which skills they exercise and clusters similar failure patterns, building a consensus signal about when a skill fails and why; (3) Skill Refinement—identified failure patterns trigger skill updates (modified prompts, adjusted tool bindings, different orchestration patterns) which are tested and validated before being pushed back to the shared library. The key mechanism is a probabilistic skill impact model that estimates whether a skill change will improve outcomes across the distribution of use-cases users actually exercise, not just in isolated test scenarios. Updated skills are staged for gradual rollout, measuring improvement metrics (task success rate, error reduction, user satisfaction) before full deployment. This approach avoids the brittleness of one-size-fits-all skill definitions by learning skill behavior from real-world usage distributions.
Production Impact
For teams running multi-user agent systems, SkillClaw eliminates the problem of repeatedly fixing the same failure modes across independent deployments—you gain a virtuous cycle where your agent library improves automatically as users interact with it. In practice, this means your skill library becomes more robust to edge cases, tool failures, and user context variations without requiring retraining or human intervention for each fix. The production pipeline changes significantly: instead of deploying frozen skill libraries and accepting static performance, you maintain an append-only trajectory store, run aggregation jobs (likely hourly or daily), and perform automated skill refinement with A/B testing before rollout. Trade-offs include increased operational complexity (you need infrastructure to collect, aggregate, and validate trajectory data safely), potential privacy concerns (trajectories contain user interactions and may require anonymization), and latency considerations (skill updates propagate asynchronously, so some users may run on older skill versions during rollout). For large-scale systems (1000+ concurrent users), the trajectory collection and aggregation cost is modest compared to LLM inference, but you must invest in robust monitoring to detect when skill updates degrade performance on specific user cohorts.
Limitations and When Not to Use This
SkillClaw assumes sufficient trajectory volume to reliably identify failure patterns, which may not hold for rare tasks or small user bases—cold-start problems require careful handling. The framework relies on explicit success/failure signals in trajectories, but many real-world tasks have ambiguous outcomes or delayed feedback (e.g., code quality, user satisfaction) that don't fit the binary success model the paper likely assumes. Skill updates based on majority consensus across users may actually degrade performance for niche use-cases that require specialized behaviors, creating a tension between collective improvement and personalization. The paper does not address adversarial scenarios: malicious users could poison the trajectory stream with false success signals, or intentionally craft trajectories that trigger incorrect skill updates. Additional work is needed on: (1) handling concept drift when user needs or tool APIs change, (2) maintaining skill stability while incorporating updates (avoiding thrashing), (3) privacy-preserving aggregation to prevent trajectory data leakage, and (4) efficient skill rollback when updates cause regressions.
Research Context
SkillClaw builds on the emerging paradigm of agentic systems that decompose complex tasks into reusable skills (prior work like OpenClaw, ToolFormer, and skill libraries in ReAct), but extends them from static to dynamic skill management. The paper connects to distributed learning and federated optimization literature, adapting techniques for learning from decentralized data to the problem of skill refinement. It advances the frontier of continual learning in production systems—similar in spirit to online learning and active learning approaches, but applied specifically to the skill library layer of agent stacks. This work opens a research direction toward truly self-improving agent systems that compound knowledge across users, with potential impact on multi-tenant AI platforms, cloud-hosted reasoning systems, and federated LLM applications where skill improvement across users is a competitive advantage.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::