Multi-Agent Orchestration Framework for long-running autonomous AI agents

OpenClaw Harness is a production-grade framework for orchestrating multi-agent AI workflows. Built on research about long-running autonomous agents, it provides:

• Auto-planning — Decompose goals into tasks using rules, LLM, or hybrid methods
• Multi-Agent orchestration — Sequential, parallel, or DAG-based execution
• Tool registry — 11 built-in tools (shell, file, Python, grep, search, etc.)
• HITL approval gates — Human-in-the-loop task approval with risk levels
• State machines — Validated state transitions for sessions, tasks, and approvals
• RAG / Vector memory — Zero-dependency TF-IDF semantic search and knowledge injection
• Multi-provider LLM — DashScope, OpenAI, Anthropic, any OpenAI-compatible endpoint
• Rate limiting — Token bucket RPM/TPM limits with 429 auto-retry
• Cost monitoring — Multi-model pricing, budget alerts, usage tracking
• Docker sandbox — Isolated code execution with resource limits
• Web dashboard — Flask-based UI with real-time state, auth, and trace export
• Structured config — Unified YAML/JSON/ENV configuration with priority resolution

pip install -i https://pypi.cnb.cool/ifree/harness/-/packages/simple openclaw-harness

git clone https://cnb.cool/ifree/harness.git
cd harness
pip install -e .

# Initialize a project
harness init ./my-project "Build a web scraper with data analysis"

# Auto-generate a task plan
harness plan ./my-project --goal "Build a web scraper" --method hybrid

# Check project status
harness status ./my-project

# Launch the dashboard
harness dashboard ./my-project --port 5001

from harness.scripts.multi_orchestrator import MultiHarnessOrchestrator, AgentRole
from harness.scripts.harness_config import HarnessConfig

# Define your team
roles = [
    AgentRole(name="Researcher", goal="需求分析", expertise=["市场研究", "竞品分析"]),
    AgentRole(name="Architect", goal="系统设计", expertise=["架构设计", "技术选型"]),
    AgentRole(name="Developer", goal="代码实现", expertise=["Python", "React"]),
    AgentRole(name="Tester", goal="质量保障", expertise=["测试", "代码审查"]),
]

# Create orchestrator
config = HarnessConfig.default()
orchestrator = MultiHarnessOrchestrator.from_config(config, roles=roles)

# Run
result = await orchestrator.run(goal="Build a REST API")

harness/
├── harness/scripts/
│   ├── orchestrator.py         # Single-agent orchestrator
│   ├── multi_orchestrator.py   # Multi-agent orchestrator
│   ├── auto_planner.py         # Auto task decomposition
│   ├── tool_registry.py        # Tool registry (11 built-in tools)
│   ├── group_chat.py           # Multi-role group chat
│   ├── rate_limiter.py         # Rate limiting (token bucket + 429 retry)
│   ├── llm_providers.py        # Multi-provider LLM abstraction
│   ├── vector_memory.py        # Vector memory / RAG (TF-IDF)
│   ├── workflow.py             # State machine + DAG workflow graphs
│   ├── dependency_manager.py   # Dependency management / critical path
│   ├── cost_monitor.py         # Cost monitoring + budget alerts
│   ├── docker_sandbox.py       # Docker sandbox execution
│   ├── dashboard.py            # Flask web dashboard (with auth)
│   ├── harness_config.py       # Unified configuration management
│   ├── tracer.py               # Distributed tracing
│   └── file_utils.py           # Atomic JSON writes
└── harness/cli/
    └── __init__.py             # CLI entry point

from harness.scripts.llm_providers import ProviderRegistry

# Auto-detect from environment
provider = ProviderRegistry.auto_detect()

# Explicit provider
from harness.scripts.llm_providers import DashScopeProvider, OpenAIProvider
provider = DashScopeProvider(api_key="sk-...", model="qwen3.6-plus")

from harness.scripts.rate_limiter import RateLimiter

limiter = RateLimiter(rpm=60, tpm=100_000, max_retries=3)

# Use with LLM client
client = LLMClient(provider=provider, rate_limiter=limiter)

from harness.scripts.workflow import WorkflowGraph, create_dag_workflow

graph = create_dag_workflow(
    nodes=["research", "design", "implement", "test"],
    edges=[
        ("research", "design"),
        ("design", "implement"),
        ("implement", "test"),
    ],
)

Create harness-config.yaml:

project:
  name: my-api
  goal: Build a REST API

llm:
  provider: dashscope
  model: qwen3.6-plus

orchestrator:
  execution_mode: hybrid
  max_concurrent: 4

rate_limit:
  enabled: true
  rpm: 60
  tpm: 100000

cost:
  daily_budget_usd: 10.0
  alert_threshold_pct: 80

Load it:

from harness.scripts.harness_config import HarnessConfig
config = HarnessConfig.load("harness-config.yaml")

pip install -e ".[dev]"
pytest tests/ -q

1262+ tests, 92% coverage.

MIT

• Source: https://cnb.cool/ifree/harness
• Docs: https://cnb.cool/ifree/harness
• Issues: https://cnb.cool/ifree/harness/-/issues
• Community: https://discord.com/invite/clawd