distributed-agent-runtime

Runtime for coordinating multiple autonomous LLM agents across distributed tasks. Battle-tested on real bug bounty pipelines (OpenClaw project) with findings up to Critical / $10k.

Architecture

┌────────────────────────────────────────────────────────────┐
│                    Task Dispatcher                          │
│         (receives goal → decomposes → assigns)             │
├──────────────┬─────────────────┬────────────────────────── ┤
│  Strategic   │  Tactical       │  Operational              │
│  Agent       │  Agent          │  Agents (N)               │
│  (80B model) │  (14B model)    │  (7B models)              │
│  Planning    │  Coordination   │  Execution                │
├──────────────┴─────────────────┴───────────────────────────┤
│                   Tool Registry                             │
│  web_search │ code_exec │ file_ops │ api_calls │ custom    │
├────────────────────────────────────────────────────────────┤
│                   Memory Layer                             │
│   Short-term (context) │ Long-term (vector DB) │ Shared   │
└────────────────────────────────────────────────────────────┘

Agent Roles

Role	Model	Responsibility
Strategist	80B (Qwen2.5/DeepSeek)	Goal decomposition, quality control
Coordinator	14B (DeepSeek-R1)	Task assignment, progress tracking
Executor	7B (Mistral)	Tool calls, data collection
Critic	14B	Output validation, retry logic

ACP Protocol

Custom Agent Communication Protocol for structured inter-agent messaging:

{
  "protocol": "acp/1.0",
  "from": "strategist",
  "to": "executor_01",
  "type": "TASK_ASSIGN",
  "task_id": "recon_001",
  "payload": {
    "objective": "...",
    "tools_allowed": ["web_search", "api_call"],
    "timeout": 300,
    "quality_threshold": 0.85
  }
}

Quick Start

from distributed_agent_runtime import AgentRuntime, Task

runtime = AgentRuntime(
    strategist_model="qwen2.5:72b",
    coordinator_model="deepseek-r1:14b",
    executor_model="mistral:7b",
    max_parallel_agents=8,
)

task = Task(
    goal="Research target and generate comprehensive report",
    tools=["web_search", "code_exec"],
    max_iterations=50,
)

result = runtime.run(task)

OpenClaw Integration

This runtime powers OpenClaw — an autonomous bug bounty pipeline:

• 80B model handles reconnaissance strategy
• 14B model coordinates scanning phases
• Multiple 7B agents execute parallel scans
• Findings validated up to Critical severity / $10k

Directory Structure

distributed-agent-runtime/
├── runtime/
│   ├── dispatcher.py
│   ├── agents/
│   │   ├── strategist.py
│   │   ├── coordinator.py
│   │   └── executor.py
│   ├── protocols/
│   │   └── acp.py
│   └── memory/
│       ├── short_term.py
│       └── long_term.py
├── tools/
├── tests/
└── examples/

License

MIT — Léo Camus / NextGen Labs