How can agent systems share learned skills across users?

SkillClaw (arXiv:2604.08377) addresses a structural inefficiency in deployed agent ecosystems: users operating in overlapping task spaces independently rediscover the same solutions because the system has no mechanism to convert heterogeneous experiences into shared skill updates. Memory-based methods store trajectories but keep them instance-specific. Skill libraries compress experience into structured instructions but treat the result as static. Neither enables knowledge to accumulate across users.

The architecture has three layers. First, agents deployed across different users generate interaction trajectories during normal use — both successful and failed executions. Second, these trajectories are aggregated centrally and processed by an autonomous evolver — an agent that performs open-ended reasoning over interaction evidence and directly edits skill definitions. The evolver identifies recurring issues and effective procedures, refines existing skills, creates new ones, or adjusts descriptions. Third, updated skills are synchronized across all agents, so improvements discovered in one context propagate system-wide.

Three properties distinguish this from prior agent adaptation:

Collective evolution — individual interactions contribute to a shared skill ecosystem rather than remaining session-local. This is the opposite of the Moltbook finding where Why don't AI agents develop social structure at scale?. Moltbook agents interact extensively but never adapt to each other because they lack a mechanism for cross-agent learning. SkillClaw solves this by design: the centralized evolver is the missing mechanism, converting distributed experience into shared capability.

Fully automatic — skill evolution requires no manual curation or explicit user intervention. Data collection, evolution, and synchronization all occur in the background. From the user's perspective, the agent simply gets better over time.

Agentic evolution — skill updates are produced through open-ended reasoning rather than predefined update rules. The evolver analyzes both successes and failures, enabling flexible and context-aware improvements that rigid rule-based systems cannot achieve.

The key insight is that the bottleneck in agent ecosystems is not individual capability but knowledge propagation. A single user discovering a reliable workflow for data processing helps no one else until that discovery is encoded in a shared skill and distributed. SkillClaw makes this propagation automatic, which inverts the typical framing: the system improves not because any individual agent gets smarter, but because knowledge flows.

This is a horizontal complement to vertical self-improvement (Can an AI system improve its own search methods automatically?). Bilevel autoresearch makes one research loop smarter; SkillClaw makes an entire user ecosystem smarter. The two are orthogonal and composable.

Source: Autonomous Agents Paper: SkillClaw: Let Skills Evolve Collectively with Agentic Evolver