The missing runtime for Agent Skills — health monitoring, self-improvement, and dependency management for any AI agent framework.
Your agent skills silently break. Agent Skill Bus detects it, diagnoses the root cause, and fixes it automatically.
Built by 合同会社みやび (LLC Miyabi) — Running 42 AI agents in production daily.
Looking for the full ecosystem? This repo is the core runtime. For 110+ production-ready skills, marketplace, and the complete Miyabi Agent Society platform, visit agentskills.bath.me.
Agent Skill Bus is a framework-agnostic runtime for AI agent skill health — orchestrating, monitoring, and self-improving agent skills across any framework. Think of it as the operational backbone that keeps your agent skills healthy over weeks and months, not just during a single run. It consists of three integrated modules:
| Module | Purpose | Standalone? |
|---|---|---|
| Prompt Request Bus | DAG-based task queue with dependency resolution & file locking | ✅ Yes |
| Self-Improving Skills | Automatic skill quality monitoring & repair loop | ✅ Yes |
| Knowledge Watcher | External change detection → automatic improvement triggers | ✅ Yes |
They work independently, but together they form a closed-loop self-improving agent system:
External Changes ──→ Knowledge Watcher ──→ Prompt Request Bus ──→ Execute
↑ │
│ ↓
Self-Improving ←── Skill Runs Log
Skills
Most agent frameworks handle execution (LangGraph, CrewAI, AutoGen). None handle operational health:
- 🔴 Skills silently degrade — An API changes, a model updates, auth expires. Nobody notices until it fails in production.
- 🔴 Tasks collide — Two agents edit the same file simultaneously. Data corruption.
- 🔴 No dependency management — Complex tasks need A→B→C ordering. Most systems just run everything in parallel.
- 🔴 No learning loop — Failures repeat because there's no feedback mechanism.
Agent Skill Bus solves all four.
# One command to set up everything
npx agent-skill-bus init
# Log a skill execution
npx agent-skill-bus record-run --agent my-agent --skill api-caller --task "fetch data" --result success --score 1.0
# Check what needs attention
npx agent-skill-bus flagged
# Queue a task
npx agent-skill-bus enqueue --source human --priority high --agent dev --task "Fix auth bug"
# See what's ready to dispatch
npx agent-skill-bus dispatchAdd this to your AGENTS.md:
After completing any task, log the result:
npx agent-skill-bus record-run --agent claude --skill <skill-name> --task "<task>" --result <success|fail|partial> --score <0.0-1.0>
That's it. The self-improving loop runs automatically.
Get a real-time overview of all your agent skills with a single command:
npx agent-skill-bus dashboard╔══════════════════════════════════════════════════════════╗
║ 🚌 Agent Skill Bus Dashboard ║
╚══════════════════════════════════════════════════════════╝
📊 Queue: 3 queued │ 1 running │ 12 completed │ 0 failed
Status Skill Score Trend Health
─────────────────────────────────────────────────────────
● ALERT api-caller 0.42 ↓ ██░░░░░░░░░░
● OK code-review 0.95 ↑ ███████████░
● OK deploy-pipeline 0.88 ─ ██████████░░
⚠ Flagged Skills:
api-caller — score_drop: dropped from 0.91 to 0.42 (drift: -53.8%)
Options: --days N (default: 7), --no-color for CI/piping.
A JSONL-based task queue with:
- DAG dependency resolution — Tasks specify
dependsOnother tasks. Automatic topological execution. - File-level locking — Prevent two agents from editing the same file. TTL-based deadlock prevention.
- Priority routing —
critical > high > medium > low. Critical tasks bypass the queue. - Multi-source ingestion — Human commands, cron jobs, GitHub webhooks, internal triggers all use the same format.
- Deduplication — Same task won't be queued twice.
{
"id": "pr-001",
"ts": "2026-03-18T08:00:00Z",
"source": "human",
"priority": "high",
"agent": "dev-agent",
"task": "Fix authentication bug in auth.ts",
"status": "queued",
"dependsOn": [],
"affectedFiles": ["myapp:src/auth.ts"],
"dagId": null
}A 7-step quality loop inspired by Cognee's self-improving agents:
OBSERVE → ANALYZE → DIAGNOSE → PROPOSE → EVALUATE → APPLY → RECORD
- Automatic failure detection — Score drops, trend analysis, consecutive failure alerts.
- LLM-powered diagnosis — Reads the failing skill + error logs, identifies root cause.
- Safe auto-repair — Low-risk fixes applied automatically. High-risk changes need human approval.
- Drift detection — Catches silent degradation (score drops >15% week-over-week).
Monitors external changes and triggers improvement requests:
- Tier 1 (every check): Dependency versions, API changes, config drift
- Tier 2 (daily): Community patterns, user feedback, platform changes
- Tier 3 (weekly): Industry trends, competitor releases, best practice updates
When a change is detected:
- Assess impact on existing skills
- Generate a Prompt Request with severity rating
- Route to Self-Improving Skills or human reviewer
| Guide | Description |
|---|---|
| Architecture Deep Dive | System design, JSONL data layer, DAG scheduling, file locking |
| Self-Improving Skills | The 7-step quality loop, drift detection, auto-repair |
| Knowledge Watcher | Three-tier monitoring, change detection, impact assessment |
| Integration Guide | Claude Code, Codex, LangGraph, CrewAI, CI/CD setup |
| Framework Comparison | Feature matrix vs. LangGraph, CrewAI, AutoGen, Mastra, VoltAgent |
┌─────────────────────────────────────────────────────┐
│ Agent Skill Bus │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌───────────┐ │
│ │ Knowledge │ │ Prompt │ │ Self- │ │
│ │ Watcher │──│ Request │──│ Improving │ │
│ │ (detect) │ │ Bus (route)│ │ (repair) │ │
│ └──────────────┘ └──────────────┘ └───────────┘ │
│ │ │ │ │
│ └──────────────────┼────────────────┘ │
│ │ │
│ ┌────────────┴────────────┐ │
│ │ JSONL Data Layer │ │
│ │ • queue.jsonl │ │
│ │ • skill-runs.jsonl │ │
│ │ • knowledge-diffs.jsonl│ │
│ │ • active-locks.jsonl │ │
│ └─────────────────────────┘ │
└─────────────────────────────────────────────────────┘
│
▼
Your Agent Framework
(OpenClaw, LangGraph, CrewAI,
Claude Code, Codex, custom...)
Detailed comparison: See docs/comparison.md for a full feature matrix vs. LangGraph, CrewAI, AutoGen, Mastra, and VoltAgent.
Agent Skill Bus is framework-agnostic. It uses plain JSONL files as the data layer — no databases, no message brokers, no vendor lock-in.
| Framework | Integration | Notes |
|---|---|---|
| OpenClaw | Native | Built and battle-tested here |
| Claude Code | Via Skills | Drop into .claude/skills/ |
| Codex | Via Skills | Drop into .codex/skills/ |
| LangGraph | Via tool calls | Read/write JSONL in tool functions |
| CrewAI | Via tool calls | Same approach |
| Custom | Direct file I/O | It's just JSONL files |
This framework runs in production at 合同会社みやび:
- 42 AI agents coordinated daily
- 27 tasks/day average throughput
- 44 cron jobs feeding the bus
- 7-minute security incident response (fastest recorded)
- 57% reduction in skill failures after enabling self-improvement loop
Agent Skill Bus is designed to work with the Agent Skills Standard — the open specification for portable AI agent skills. Each skill follows the SKILL.md format and can be used across Claude Code, Codex, OpenClaw, and any compatible agent runtime.
your-project/
├── skills/
│ ├── api-caller/
│ │ └── SKILL.md # Skill definition (Agent Skills Standard)
│ ├── code-reviewer/
│ │ └── SKILL.md
│ └── deploy-pipeline/
│ └── SKILL.md
├── .agent-skill-bus/
│ ├── queue.jsonl # Task queue
│ ├── skill-runs.jsonl # Execution history
│ ├── knowledge-diffs.jsonl # Detected changes
│ └── active-locks.jsonl # File locks
The Self-Improving Skills module reads SKILL.md files to understand what each skill does, diagnoses failures in context, and can propose targeted fixes — all without any skill-specific configuration.
Issues and PRs welcome. See CONTRIBUTING.md.
MIT — See LICENSE.
Building the future of AI agent operations.
AIエージェントシステム向けの自己改善型タスクオーケストレーションフレームワークです。複数のAIエージェントが協調してタスクを実行する際に必要な、タスクキュー・スキル品質監視・ナレッジ管理の3つの機能を提供します。
Claude Code、Codex、LangGraph、CrewAI、AutoGenなど、あらゆるLLMエージェントと組み合わせて使用できます。依存関係ゼロ、MITライセンス。
npm install agent-skill-bus# プロジェクトを初期化
npx skill-bus init
# タスクキューを起動
npx skill-bus run| モジュール | 名称 | 説明 |
|---|---|---|
| Prompt Request Bus | DAGタスクキュー | 依存関係を考慮したタスクの並列・逐次実行 |
| Self-Improving Skills | スキル品質モニタリング | 各スキルの成功率・失敗パターンを追跡し、自動改善サイクルを実現 |
| Knowledge Watcher | ナレッジウォッチャー | コードベースやドキュメントの変更を監視し、エージェントのコンテキストを最新に維持 |
MIT License — 合同会社みやび（代表: 林 駿甫）
完全版はこちら → このリポジトリは Agent Skill Bus のコアランタイムです。110以上の本番スキル、マーケットプレイス、Miyabi Agent Society の全機能は agentskills.bath.me でご覧いただけます。
Full version here → This repo is the core runtime only. For 110+ production-ready skills, the marketplace, and the complete Miyabi Agent Society platform, head to agentskills.bath.me.