Source-linked AI summary
Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification for Coordination Engineering
Xinyu Zhang, Zhicheng Dou, Deyang Li, Jianjun Tao, Shuo Cheng, Ruifeng Shi, Fangchao Liu, Enrui Hu, Yangkai Ding, Hongbo Wang, Qi Ye, Xuefeng Jin, Zhangchun Zhao
TL;DR
Multi-agent coordination protocols are often locked in framework-specific code or configurations, while successful collaboration experience can be lost between sessions. Swarm Skills provides a portable specification and companion self-evolution algorithm for creating and refining coordination assets. The paper presents architectural analysis and a qualitative JiuwenSwarm case study as evidence of zero-adapter portability, while noting that broad quantitative conformance and performance evaluation remain open.
Problem
Multi-agent coordination protocols remain framework-bound, and coordination decisions can be discarded with execution traces instead of being reused or autonomously improved.
Source: §1 · pp. 1–2Method
Swarm Skills extends Anthropic Skills with a portable multi-agent asset structure and self-evolution semantics, paired with an algorithm that creates skills from successful trajectories and patches existing ones using friction patterns and multidimensional scores.
Source: §1 · p. 2Results
Takeaways & Limitations
The paper concludes that coordination protocols can be represented as portable assets separated from runtimes and structured for continuous refinement.
Source: §4 · p. 4, §5 · p. 5Large-scale quantitative conformance tests across diverse Host Agents and standardized benchmarks for measuring coordination gains are absent; effectiveness on hosts lacking recursive file reading or dynamic sub-agent spawning remains unevaluated.
Source: §4 · p. 4
Abstract
from arXiv · showhide
As artificial intelligence engineering paradigms shift from single-agent Prompt and Context Engineering toward multi-agent \textbf{Coordination Engineering}, the ability to codify and systematically improve how multiple agents collaborate has emerged as a critical bottleneck. While single-agent skills can now be distributed as portable assets, multi-agent coordination protocols remain locked within framework-internal code or static configurations, preventing them from being shared across systems or autonomously improved over time. We propose \textbf{Swarm Skills}, a portable specification that extends the Anthropic Skills standard with multi-agent semantics. Swarm Skills turns multi-agent workflows into first-class, distributable assets that consist of roles, workflows, execution bounds, and a built-in semantic structure for self-evolution. To operationalize the specification's evolving nature, we present a companion self-evolution algorithm that automatically distills successful execution trajectories into new Swarm Skills and continuously patches existing ones based on multi-dimensional scoring (Effectiveness, Utilization, and Freshness), eliminating the need for human-in-the-loop oversight during the refinement process. Through an architectural compatibility analysis and a comprehensive qualitative case study using the open-source JiuwenSwarm reference implementation, we demonstrate how Swarm Skills achieves zero-adapter cross-agent portability via progressive disclosure, enabling agent teams to self-evolve their coordination strategies without framework lock-in.
1 Introduction
Multi-agent coordination experience is often lost after a session and remains tied to particular frameworks. Swarm Skills makes coordination a portable, self-evolving asset, paired with an algorithm for refining it and a qualitative portability case study.
1 Introduction
- The paper frames coordination engineering as a response to single-agent limitations on complex tasks, where orchestration, task allocation, communication, and isolation become central.It positions reusable multi-agent coordination experience as a missing component alongside portable and evolvable single-agent skills.Source: §1 · p. 1
- Coordination decisions are often discarded with execution traces, leaving similar tasks to be planned from scratch and protocols unable to travel across frameworks or improve natively.The missing capability is described as coordination experience sedimentation: capturing, reusing, and iteratively improving successful collaboration patterns.Source: §1 · pp. 1–2
- Swarm Skills extends Anthropic Skills with a distributable, runtime-separated format for multi-agent coordination and built-in semantics for accumulating operational feedback.Its five components are frontmatter, roles, workflow, execution bounds, and dependencies.Source: §1 · p. 2
- The companion algorithm distills successful trajectories into new skills and patches existing ones using friction patterns and Effectiveness, Utilization, and Freshness scoring.Automated simplification, culling, rebuilding, and rollback routines are used to limit low-quality patches, without human approval gates.Source: §1 · p. 2
- Architectural analysis and a qualitative six-role travel-planning case study using JiuwenSwarm are presented as evidence of zero-adapter cross-agent portability.The figure contrasts framework-bound coordination and discarded traces with a portable asset that accumulates coordination experience across sessions.Source: Figure 1 · p. 2, §1 · p. 3
2 Related Work
Prior work establishes portable and evolvable single-agent skills and explores multi-agent roles, topologies, and workflow search. Its coordination protocols generally remain internal to frameworks or optimization procedures; Swarm Skills instead proposes a portable description format.
2.1 Skill Abstractions and Single-Agent Self-Evolution
- Voyager [10], SAGE, ExpeL [15], and EvoSkills develop reusable or self-evolving single-agent skills, while this work extends Anthropic Skills portability to multi-agent settings.These systems establish skills as portable, evolvable assets, but are characterized here as constrained to the single-agent paradigm.Source: §2.1 · p. 3
2.2 Multi-Agent Topologies and Workflow Optimization
- CAMEL [3], Generative Agents [6], ChatDev [7], AgentVerse [1], DyLAN [5], GPTSwarm, AFlow [14], and EvoAgent [13] explore agent interaction, topology optimization, or workflow search.The resulting coordination protocols are described as internal data structures rather than portable descriptions for cross-framework distribution.Source: §2.2 · p. 3
2.3 Existing Multi-Agent Descriptive Formats
- CrewAI templates, MetaGPT SOPs [2], and AutoGen GroupChat configurations are framework-private formats without explicit built-in self-evolution semantics.The specification addresses this by decoupling coordination descriptions from execution runtimes.Source: §2.3 · p. 3
2.4 Engineering Paradigms for LLM Agents
- The paper places Swarm Skills within Coordination Engineering, which it describes as orchestrating multi-agent systems through team formation, communication protocols, and fault recovery.It identifies coordination experience sedimentation as the particular challenge Swarm Skills addresses.Source: §2.4 · p. 4
3 Motivation: Multi-Agent Skill Demand in the Wild
A repository measurement study finds that users already express multi-role collaboration needs through single-agent skill files. The authors interpret these workarounds and adoption patterns as evidence of demand for shareable multi-agent coordination protocols.
3 Motivation: Multi-Agent Skill Demand in the Wild
- 33 search queries across 9 Anthropic Skills repositories, crawled in April 2026, were used to measure community demand for multi-agent capabilities.The authors characterize the observed patterns as revealed preference.Source: §3 · p. 4
- Examples of team-style skills include c-level-advisor with 9 roles, ra-qm-team with 13 personas, and engineering-team with 14 roles.Critique skills had more than 77.8K installs, alongside adoption of parallel-debugging and adversarial-reviewer skills.Source: §3 · p. 4
- The findings indicate that users work around single-agent formats to express multi-role collaboration needs, supporting demand for shareable coordination protocols.The authors identify this as a pressing demand within Coordination Engineering.Source: §3 · p. 4
4 Discussion, Limitations, and Future Work
The paper reports feasibility for zero-adapter deployment and autonomous refinement, while identifying evaluation and broader coordination challenges that remain unresolved.
4 Discussion, Limitations, and Future Work
- Architectural analysis and a JiuwenSwarm case study demonstrate the feasibility of zero-adapter deployment and autonomous refinement.The paper also states that several challenges remain.Source: §4 · p. 4
- Large-scale quantitative conformance testing across diverse Host Agents is absent, and effectiveness on hosts lacking recursive file reading or dynamic sub-agent spawning remains unevaluated.The paper also notes that absent standardized coordination benchmarks make performance gains difficult to quantify across domains.Source: §4 · p. 4
- Future work targets empirical tests across commercial and open-source frameworks, governance for first-run lock-in, and broader challenges in fault tolerance, message routing, and inter-agent isolation.The authors call for continued community collaboration on these broader challenges.Source: §4 · p. 4
5 The Swarm Skills Specification
Swarm Skills specifies coordination as a portable description format rather than a runtime. It combines orchestration metadata and agent-facing instructions with an evolution record, while leaving execution mechanics to the Host Agent.
5 The Swarm Skills Specification
- Swarm Skills separates coordination descriptions from framework-specific execution, allowing a compliant Host Agent to run workflows through its native toolchain.The specification also embeds self-evolving semantics in the asset structure.Source: §5 · p. 5
5.1 Specification Architecture and the ”Five Components”
- A Swarm Skill is a directory with five author-provided components and an automatically generated evolution artifact, loaded progressively as needed.The Host Agent initially loads only minimum metadata before fetching detailed instructions.Source: §5.1 · p. 5
5.1.1 Frontmatter Design for Mechanical Consumption
- The YAML frontmatter supplies mechanical orchestration metadata, while additionalProperties lets unsupported Host Agents ignore multi-agent fields and treat the asset as a standard single-agent skill.The kind discriminator identifies a Swarm Skill, and roles[] specifies participant environment configuration.Source: §5.1.1 · p. 6
5.1.2 Natural Language Body for Agentic Consumption
- The natural-language body gives the Leader workflow overview and trigger conditions; roles/, workflow.md, and bind.md provide personas, task dependencies, and execution boundaries.Boundaries can include message-turn limits, token budgets, and quality gates.Source: §5.1.2 · p. 6
5.2 Embedded Self-Evolving Semantics
- The required evolutions.json stores discrete workflow or role patches, making accumulated coordination improvements portable alongside the base protocol.Evolution Records include friction context, a change directive, and Effectiveness, Utilization, and Freshness metrics.Source: §5.2 · pp. 6–7
5.3 Design Trade-offs and the Boundary of Specification
- The specification leaves message passing, task queues, and process isolation to the Host Agent, declaring collaborators, task dependencies, and execution constraints instead.This restraint is presented as enabling portability across different architectures and zero-adapter compatibility with the existing ecosystem.Source: §5.3 · p. 7
6 A Self-Evolution Algorithm for Swarm Skills
The companion algorithm distills coordination trajectories into skills, loads them progressively for use, and appends friction-driven patches. Scoring and governance routines curate accumulated records; the algorithm is a reference mechanism, not a requirement of the specification.
6 A Self-Evolution Algorithm for Swarm Skills
- The algorithm provides a reference mechanism to extract coordination protocols from trajectories, patch workflows for runtime friction, and curate improvements through multidimensional scoring.The specification does not require this particular algorithm; conforming Host Agents may implement alternatives.Source: §6 · p. 7
6.1 The Three-Stage Lifecycle
- CREATE, USE, and PATCH form a continuous lifecycle for acquiring and refining a Swarm Skill.Figure 3 depicts the loop and its governance actions.Source: §6.1 · p. 7, Figure 3 · p. 8
6.1.1 CREATE: Trajectory Distillation
- CREATE detects sufficiently complex collaboration, then distills a trajectory into a generalized workflow and role configurations stored in a staging directory.Adoption depends on the Host Agent’s subsequent skill selection during USE.Source: §6.1.1 · pp. 7–8
6.1.2 USE: Progressive Disclosure and Agentic Routing
- USE exposes skill descriptions in the Leader’s prompt, then loads selected instructions, role, workflow, and boundary files incrementally, appending applicable Evolution Records.Descriptions are surfaced without loading the full implementation into the prompt.Source: §6.1.2 · p. 8
6.1.3 PATCH: Friction-Driven Optimization
- PATCH scans traces for implicit coordination friction and explicit improvement signals, then appends new Evolution Records without altering the base specification files.The unchanged base files mitigate the risk of catastrophic degradation from a single erroneous patch.Source: §6.1.3 · pp. 8–9
6.2 Multi-Dimensional Scoring and Automated Culling
- Each Evolution Record receives a composite score based on effectiveness, utilization, and freshness; low-scoring records are eventually ignored by the Host Agent.Effectiveness uses Bayesian smoothing; utilization decays when the Leader ignores an instruction, while freshness decays over time.Source: §6.2 · p. 9
6.3 Governance Actions: SIMPLIFY, REBUILD, and ROLLBACK
- When the evolution record array exceeds a specified capacity, SIMPLIFY prunes or merges records, REBUILD incorporates patches into the base files, and ROLLBACK restores an archived state.The previous version is archived during REBUILD, allowing recovery from a detrimental rebuild.Source: §6.3 · p. 9
6.4 The JiuwenSwarm Reference Implementation
- JiuwenSwarm implements the lifecycle and governance routines by interacting with the specification schema, without modifying the underlying execution framework.The paper presents this as demonstrating that operationalizing self-evolving semantics can itself be portable.Source: §6.4 · p. 9
7 Architectural Compatibility & Deployment
Swarm Skills use the Anthropic Skills standard’s progressive disclosure to support deployment across compatible hosts without framework-specific translation. Hosts without multi-agent capabilities can treat the asset as a standard single-agent Skill, though large-scale commercial-framework conformance testing remains future work.
7.1 Compatibility via Progressive Disclosure
- Anthropic-compatible hosts can deploy a Swarm Skill by copying its directory; native file reading loads roles and workflow, while native sub-agent spawning instantiates teammates.Source: §7.1 · p. 10
- Natural-language workflows in Markdown avoid proprietary execution-graph parsing, unlike framework-specific configurations that require translation adapters.Source: §7.1 · p. 10
7.2 Graceful Degradation
- Hosts without multi-agent capabilities can ignore the swarm discriminator and roles array, then execute the workflow sequentially as a standard single-agent Skill.Source: §7.2 · p. 10
- Large-scale empirical conformance testing across multiple commercial frameworks remains future work, so portability is supported by architectural analysis rather than that testing.Source: §7.2 · p. 10
8 Qualitative Case Study: The travel-planning-swarm Skill
A JiuwenSwarm walkthrough of a constrained family trip illustrates expert coordination and post-run skill evolution. The team reconciles a scheduling conflict, then refines its roles and role-specific behaviors.
8.1 Design Rationale and Initial Setup
- The five-role team divided transport, accommodation, attractions, plan synthesis, and budget review with social-media drafting for a family trip capped at 20,000 RMB.Source: §8.1 · p. 11
8.2 System Walkthrough: Autonomous Conflict Resolution
- Experts worked concurrently, cross-checking proposals and resolving temporal or spatial conflicts before sending reconciled plans to the synthesizer.Source: §8.2 · p. 11, Figure 4 · p. 11
- After detecting an impossible arrival-and-itinerary overlap, Transportation and Attraction agents negotiated a morning flight, preserving afternoon sightseeing within the 20,000 RMB limit.Source: §8.2 · pp. 11–12
8.3 Demonstrating the Evolution Experience
- Post-run analysis flagged the Budget Reviewer’s coupled auditing and copywriting duties; after user approval, REBUILD added a sixth Copywriting Expert and routed copywriting in parallel.Source: §8.3 · p. 12
- Transportation and Attraction negotiation tactics, including prioritizing infant comfort over minor cost savings, were distilled into behavioral patches for their role files.Source: §8.3 · p. 12
- The case presents repeated use as strengthening team capability, while JiuwenSwarm interactive mode optionally adds human review for structural changes.Source: §8.3 · p. 12