52e3470e06
This commit introduces a multi-agent Swarm Engine to PicoClaw, enabling complex task orchestration through autonomous collaboration and long-term semantic memory. Key Features: - Actor Model Architecture: Every agent (node) runs in an isolated goroutine, communicating via a lightweight internal Event Bus. - Hierarchical Delegation: Manager nodes can dynamically spawn specialized worker nodes (Researchers, Analysts, etc.) to perform sub-tasks in parallel. - Hybrid Memory System: Combined SQLite for state persistence and Chromem-go for a persistent, cross-swarm Vector Knowledge Base. - Progressive Summarization: Implemented an intelligent memory pruning mechanism that summarizes older context before truncation to preserve findings without exceeding token limits. - Role-Based Access Control (RBAC): Configurable tool policies per role (e.g., Researchers can browse but not execute shell commands). Performance & Scaling: - Low Footprint: Idle memory usage is ~7-8 MB RSS. - Efficient Scaling: During a stress test with 10 concurrent agents performing intensive research, memory usage peaked at only 27 MB. - Per-Node Cost: Each active agent consumes approximately 0.6 MB to 2 MB of physical RAM depending on conversation length and summarization state. - Stability: Successfully handled 10 parallel LLM requests with zero race conditions, showcasing the robustness of the Go-based Actor model. Integration: - Native CLI Support: Added /swarm spawn, /swarm list, and /swarm status commands. - Mermaid Visualization: Support for /swarm viz <id> to generate organizational charts of active agent hierarchies. - Configurable: Roles, models, and security policies are fully customizable via config.json. |
||
|---|---|---|
| .. | ||
| adapters | ||
| bus | ||
| config | ||
| core | ||
| memory | ||
| prompt | ||
| runtime | ||
| README.md | ||
| service.go | ||
PicoClaw Swarm Engine
This package implements the Swarm Agents architecture: a lightweight, hierarchical, multi-agent system designed for the edge.
For a detailed deep-dive into the architecture, workflow, and components, please read SWARM_ARCHITECTURE.md.
Features
- Hierarchical Swarms: Manager nodes can delegate tasks to specialized workers (Researchers, Writers, etc.).
- Vector Memory: Integrated
Chromem-govector database for long-term semantic memory. - Resilient: Built-in retry mechanisms, context pruning, and crash recovery via SQLite checkpoints.
- Secure: Configurable Role-Based Access Control (RBAC) for tools.
- Lightweight: Built on Go Goroutines (Actor Model) and Channels, efficient enough for small VPS or local devices.
Usage
The Swarm Engine is integrated into the PicoClaw CLI.
Basic Commands
# Spawn a new swarm to perform a task
/swarm spawn "Analyze the latest trends in Rust vs Go performance"
# List active swarms
/swarm list
# Check detailed status of a specific swarm (including node progress)
/swarm status <swarm_id>
# Visualize the swarm topology (Mermaid diagram)
/swarm viz <swarm_id>
# Stop a running swarm
/swarm stop <swarm_id>
Configuration
You can configure the Swarm behavior, Roles, and Policies in your config.json.
{
"swarm": {
"limits": {
"max_nodes": 10,
"global_timeout": 600000000000,
"max_iterations": 15
},
"roles": {
"Manager": {
"description": "Orchestrates tasks",
"system_prompt": "You are a Manager...",
"tools": ["delegate_task", "save_memory", "search_memory"]
}
},
"policies": [
{
"role": "Manager",
"allowed": ["*"]
}
]
}
}
Architecture Overview
graph TD
Service --> Orchestrator
Orchestrator --> Manager[Manager Node]
Manager -->|Delegates| Worker[Worker Node]
Manager <-->|Memory| VectorDB[(Vector DB)]
Manager <-->|State| SQLite[(SQLite)]