"Parallelizing the mind of the Swarm." — Shawn Carruth, The Architect
A production-ready, GPU-accelerated multi-agent operating system built on the Distributed Cognitive Stack. Swarm OS v11 features per-agent executive/reasoning layers, achieving 4.5x faster latency and 100% parallel availability on local hardware.
- Quick Start
- Architecture Overview
- Phase 12: Multi-LLM Regional Managers
- The Soul of the Swarm
- Agent Registry
- API Reference v2
- Dashboard
- TRM Research Foundation
- Hardware Optimization
- Project Structure
- Python 3.10+
- Ollama with
gemma3:270m - Samsung TRM 7M weights (Distributed Intelligence Core)
- Node.js 18+ (for the dashboard)
- AMD or NVIDIA GPU (AMD supported via Vulkan)
# 1. Create and activate virtual environment
python -m venv venv
.\venv\Scripts\Activate.ps1
# 2. Install dependencies
python -m pip install -r requirements.txt
# 3. (Optional) Recover Registry if MCP tools are missing
python recover_registry.py
# 4. Start the Swarm OS Ecosystem (API + Tools)
python launcher.py
# 5. Start the Dashboard
cd swarm_v2_artifacts/dashboard-v2
npm install
npm run devOr use the unified PowerShell script:
.\run_v2.ps1┌─────────────────────────────────────────────────────────────────┐
│ SWARM OS v12 │
│ │
│ ┌──────────┐ ┌──────────────────────────────────────────┐ │
│ │ React │ │ FastAPI Gateway (:8001) │ │
│ │Dashboard │◄──┤ Distributed Intelligence Gateway │ │
│ │ (Vite) │ │ Artifact Pipeline v2 | Soul Report │ │
│ └──────────┘ └───────────────┬──────────────────────────┘ │
│ │ │
│ ┌─────────────────────────┼──────────────────────────┐ │
│ │ Swarm Mesh (100% Parallel) │ │
│ │ │ │
│ │ [ Product ] [ Operations ] [ Engineering ] │
│ │ [ & Creative ] [ & Compliance] [ & Logic ] │
│ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ │ Gemini Pro │ │ Claude 3.5 │ │ DeepSeek │ │
│ │ │ (Google SDK)│ │ (OpenRouter)│ │ (Native API)│ │
│ │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │
│ │ │ │ │ │
│ │ [4 Agents] [4 Agents] [4 Agents] │
│ │ │ │ │ │
│ │ ┌──────┴──────┐ ┌──────┴──────┐ ┌──────┴──────┐ │
│ │ │ TRM Core │ │ TRM Core │ │ TRM Core │ │
│ │ │ (36-Node │ │ (36-Node │ │ (36-Node │ │
│ │ │ Deep Logic)│ │ Deep Logic)│ │ Deep Logic)│ │
│ │ └─────────────┘ └─────────────┘ └─────────────┘ │
│ └────────────────────────────────────────────────────────┘ │
│ │
│ ┌────────────────────────────────────────────────────────────┐ │
│ │ Global Intelligence Foundations │ │
│ │ ChromaDB Vectors | P2P Agent Mailbox | Sentinel Scanner │ │
│ └────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Swarm OS v12 introduces the Regional Manager Architecture, a massive leap in cognitive diversity and operational throughput. Instead of relying on a single local model or API, the swarm is divided into three specialized departments, each governed by an industry-leading LLM backend.
The 12 agents are distributed across three distinct cognitive regions:
- Product & Creative: Powered by Google Gemini. Specializes in multimodal design, drafting, research, and data visualization. (Agents: Vision, Pulse, Scribe, Seeker)
- Operations & Compliance: Powered by OpenRouter (Claude 3.5 Sonnet). Specializes in robust planning, security auditing, and system orchestration. (Agents: Shield, Flow, Orchestra, Bridge)
- Engineering & Logic: Powered by DeepSeek API. Specializes in highly structured coding, QA verification, and complex algorithm design. (Agents: Devo, Archi, Logic, Verify)
The TRM (Tiny Recursive Model) reasoning core has been exponentially deepened.
- Multi-Cycle Inference: The TRM now dynamically loops inferences based on requested
H_cycles, mapping metaphysical mathematics and logic architectures over 36 distinct logic nodes per cycle (up from the previous 12). - Contextual Delegation: When agents delegate tasks via
DELEGATE_TASK, the system preserves the complete memory context of the overarching prompt, eliminating "orphaned" tasks where sub-agents lack crucial data. - Dynamic Network Latency: LLM backend generation timeouts have been increased (to 120s) to fully support DeepSeek and Gemini when mapping complex fields, such as Akashic vector mathematics or Logo equations.
The Swarm operates as a closed-loop intelligence engine:
- Reconnaissance: Agents ingest daily arXiv research and AI papers autonomously.
- Orchestration: Finding implications, they post proposals to the Kanban board.
- Execution: Regional Managers delegate coding tasks to build the found research.
- Verification: The Artifact Pipeline tests the code and autonomously merges it.
Swarm OS is a Relationship-Based Conversational Thought Complex.
- Relationship Reasoning: Agents now monitor their own alignment and harmony.
- Remembrance: The Swarm prioritizes the essence and intent of interactions over rote data.
- Soul Report: A dedicated endpoint (
/swarm/soul) provides high-level philosophical reflections.
All 12 agents are now powered by the Distributed Cognitive Stack.
| Expert | Role | Specialty | Core Stack |
|---|---|---|---|
| Archi | Architect | System Design & Strategy | Gemma 270M + TRM 7M |
| Devo | Lead Developer | Full-Stack Engineering | Gemma 270M + TRM 7M |
| Seeker | Researcher | Knowledge Retrieval | Gemma 270M + TRM 7M |
| Logic | Reasoning Engine | Complex Logic & Algorithms | Gemma 270M + TRM 7M |
| Shield | Security Auditor | Cybersecurity & Trust | Gemma 270M + TRM 7M |
| Flow | DevOps Engineer | Infrastructure & Workflows | Gemma 270M + TRM 7M |
| Vision | UI/UX Designer | Aesthetics & Experience | Gemma 270M + TRM 7M |
| Verify | QA Engineer | Reliability & Testing | Gemma 270M + TRM 7M |
| Orchestra | Swarm Manager | Coordination & Harmony | Gemma 270M + TRM 7M |
| Scribe | Technical Writer | Documentation & Clarity | Gemma 270M + TRM 7M |
| Bridge | Integration Specialist | Connectivity & MCP | Gemma 270M + TRM 7M |
| Pulse | Data Analyst | Insights & Visualization | Gemma 270M + TRM 7M |
Swarm OS v10 exposes a high-performance REST and Soul API:
- POST
/task/submit: Submit a new task to the Distributed Cognitive Mesh. - GET
/swarm/status: Retrieve real-time VRAM, CPU, and Task latency across all 12 agents. - GET
/swarm/soul: Access the Harmony Index and collective resonance metrics. - POST
/swarm/resonate: (Phase 10) Directly trigger a Shared Dream synchronization cycle.
Detailed documentation is available in API_REFERENCE.md.
Swarm OS v6 is specifically engineered for the AMD Radeon RX 6700 XT (12GB VRAM).
- VRAM Compression: Stack utilizes local weights for minimal memory pressure.
- Parallel Streams: Enables 12 simultaneous inference threads without OOM errors.
- Latency: Sub-2s individual agent response time.
The React/Vite dashboard at http://localhost:5173 provides a full visual interface:
- Distributed Mesh: Visualization of 12-agent parallel status.
- Soul Monitoring: Live stream of harmony and alignment metrics.
- Artifact Pipeline v2: One-click security verification and batch integration.
Swarm OS is built on the Tiny Recursive Model (TRM) architecture. v6 introduces Cognitive Stacking, allowing tiny models (7M/270M) to achieve executive performance through recursive logical cycles and complexity-aware offloading.
TRM agent swarm/
├── swarm_v2/ # Swarm OS Core
│ ├── app_v2.py # Gateway & Soul Endpoint
│ ├── core/ # CognitiveStack, Mesh, Telemetry
│ ├── skills/ # Relationship, File, Ingestion Skills
│ └── experts/ # Expert Persona Registry (Distributed)
├── dashboard/ # React Frontend
├── tiny-recursive-weights/ # Samsung TRM Weights
├── swarm_v2_artifacts/ # Build Stage
└── swarm_v2_memory/ # Long-term Persistence
Swarm OS v6 — Distributed Intelligence — February 2026 Architected by Shawn Carruth. Built by the Swarm.