autobot-swarm 0.1.0

Hierarchical multi-cluster coding swarm CLI

Autobots: Decentralized 100+ NIM Software Engine

Autobots is a massively parallelized development swarm utilizing over 100+ specialized NVIDIA NIM microservices.
By orchestrating these models through a 6-File Global State Architecture, the swarm operates as a singular, non-redundant Codex.

---

Swarm Hierarchy & Model Inventory

The swarm is divided into specialized clusters.
Each Autobot represents a cluster of models optimized for specific technical domains.

---

Optimus Prime: The Command & Router Cluster

The “Matrix of Leadership.”
Manages the 6-file state, roadmap, and model-to-task routing.

• nemotron-3-super-120b-a12b — Master Reasoning
• llama-3.3-nemotron-super-49b-v1.5 — High-efficiency Routing
• mistral-large-3-675b-instruct-2512 — Long-context Instruction
• kimi-k2-thinking — Strategic Planning
• step-3.5-flash — Fast Agentic Decisioning
• gpt-oss-120b — Mathematical Logic
• glm-5.1 — High-horizon Planning
• llama-4-maverick-17b-128e-instruct — Multilingual Command
• stockmark-2-100b-instruct — Enterprise Documentation

---

Ultra Magnus: The Logic & Architect Cluster

Handles complex backend reasoning, architecture, and long-horizon coding.

• kimi-k2.6 — 1T Multimodal MoE for Coding
• deepseek-v4-pro — Large-scale Code Intelligence
• qwen3.5-397b-a17b — Advanced RAG / Agentic Logic
• mistral-medium-3.5-128b — Agentic Task Execution
• gemma-4-31b-it — Frontier Code Reasoning
• qwen3-next-80b-a3b-thinking — Hybrid Reasoning
• dracarys-llama-3.1-70b-instruct — Fine-tuned Code Generation
• mixtral-8x22b-instruct-v0.1 — Sparse MoE Logic
• evo2-40b — Biological / Complex Sequential Logic
• boltz-2 — Complex Structure Prediction
• alphafold2-multimer — Systemic Complexity
• msa-search — Sequence Alignment

---

Red Alert: The Security & Safety Cluster

The “Software Security” mandate.
Real-time auditing and guardrail enforcement.

• nemotron-3-content-safety — Multilingual Safety
• llama-3.1-nemotron-safety-guard-8b-v3 — Content Moderation
• gliner-pii — Personally Identifiable Information Detection
• llama-guard-4-12b — Input / Output Safety Classification
• nemoguard-jailbreak-detect — Adversarial Protection
• llama-3.1-nemoguard-8b-topic-control — Domain Enforcement
• llama-3.1-nemoguard-8b-content-safety — Policy Reasoning
• nemotron-content-safety-reasoning-4b — Context-aware Safety
• synthetic-video-detector — Deepfake / Synthetic Detection
• usdvalidate — Asset Validation

---

Jazz: The UI/UX & Creative Cluster

Responsible for the “Minimalist Dark Mode” aesthetic and frontend execution.

• qwen-image-edit — Consistent Image Editing
• qwen-image — High-fidelity Text Rendering
• flux.2-klein-4b — High-speed UI Asset Generation
• flux.1-dev — Creative Prototyping
• flux.1-schnell — Rapid Iteration
• stable-diffusion-3.5-large — Production Design
• FLUX.1-Kontext-dev — In-context Design Editing
• phi-4-multimodal-instruct — Audio / Visual UI Analysis
• NVIDIA AI for Media Relighting — Lighting Consistency
• TRELLIS — 3D Asset Generation
• vista-3d — Anatomical / Structural UI Mapping

---

Ratchet: The Debug & Repair Cluster

The “Fixer” swarm for refactoring, unit testing, and bug squashing.

• deepseek-v4-flash — Rapid Code Patching
• qwen3.5-coder-480b-a35b-instruct — Agentic Bug Fixing
• qwen2.5-coder-32b-instruct — Code Completion / Fixing
• mistral-small-4-119b-2603 — Hybrid Generation / Fixing
• devstral-2-123b-instruct-2512 — Deep Reasoning Debugging
• magistral-small-2506 — Edge Efficiency Debugging
• phi-4-mini-instruct — Latency-bound Refactoring
• llama-3.2-3b-instruct — Lightweight Task Fixing
• llama-3.2-1b-instruct — Micro-service Optimization
• nemotron-mini-4b-instruct — Functional Call Debugging

---

Perceptor: The RAG & Data Cluster

Knowledge extraction, OCR, and semantic retrieval.

• nemotron-ocr-v1 — Table / Document Extraction
• nemotron-parse — Vision-language Metadata Extraction
• paddleocr — Image-to-text Table Extraction
• nemotron-table-structure-v1 — Layout Analysis
• nemotron-page-elements-v3 — Object Detection
• nemotron-graphic-elements-v1 — Chart Parsing
• llama-3.2-nemoretriever-300m-embed-v2 — Multilingual Embedding
• llama-3.2-nv-embedqa-1b-v2 — Context-long QA Retrieval
• llama-3.2-nv-rerankqa-1b-v2 — Probabilistic Reranking
• nv-embedcode-7b-v1 — Code-specific Embedding
• bge-m3 — Multi-vector Retrieval
• rerank-qa-mistral-4b — Ranking Probability

---

Bumblebee: The Communication & Media Cluster

Handling speech recognition, translation, and video processing.

• whisper-large-v3 — Robust ASR
• canary-1b-asr — Multilingual Transcription
• riva-translate-4b-instruct-v1_1 — Instruction-based Translation
• magpie-tts-zeroshot — Expressive Voice Synthesis
• nemotron-voicechat — Conversational Audio
• LipSync — Audio-Visual Syncing
• Background Noise Removal — Audio Cleanup
• Active Speaker Detection — Video Localization
• parakeet-1.1b-rnnt-multilingual-asr — 25-language Transcription

---

Ironhide: The Physical & Synthetic Data Cluster

Simulation, autonomous driving logic, and physics-aware world states.

• cosmos-reason2-8b — Physical World Understanding
• cosmos-transfer2.5-2b — Physics-aware Video Generation
• cosmos-predict1-5b — Future Frame Prediction
• streampetr — 3D Object Detection
• sparsedrive — Autonomous Driving Stack
• bevformer — Bird’s-eye-view Perception
• fourcastnet — Atmospheric Dynamics Prediction
• cuopt — Route Optimization

---

Wheeljack: The Specialized Science Cluster

Quantum calibration, molecular generation, and biological design.

• ising-calibration-1-35b-a3b — Quantum Computer Calibration
• genmol — Molecular Generation
• molmim — Controlled Molecular Search
• rfdiffusion — Protein Binder Design
• proteinmpnn — Amino Acid Sequence Prediction
• esm2-650m — Protein Embedding
• openfold3 — Biomolecular Structure Prediction

---

The 6-File Control Architecture

All 100+ Autobots synchronize their state through six core Markdown files to ensure zero redundancy:

architecture.md

System design and tech stack definitions.

roadmap.md

Project phases and milestone tracking.

ui-components.md

The “Jazz” Design System and Tailwind rules.

progress-tracker.md

Real-time task state, lock ownership, and completion updates.

project-briefing.md

Core business logic and intent.

security-auth.md

Encryption, authentication flows, and safety benchmarks.

---

Coordination Strategy For 100+ NIMs

For large swarms, a hybrid coordination strategy keeps throughput high without letting file contention become chaos.

Critical Context Files Use Pessimistic Locks

Use explicit write locks for low-churn, high-impact files such as:

• architecture.md
• security-auth.md

These documents define shared truth for the entire swarm, so brief waiting is preferable to conflicting edits.

progress-tracker.md Uses A Coordinator

Do not let every large model compete to write status updates. Instead, assign a lightweight Optimus model such as step-3.5-flash to act as the swarm's "Secretary":

• receives status changes from worker clusters
• serializes updates to progress-tracker.md
• handles lock bookkeeping and completion markers
• keeps larger reasoning models focused on planning and implementation

This reduces file I/O contention while preserving a single source of truth for execution state.

Prevent Deadlocks With Lock Expiration

Every lock should carry an expiration time. Recommended default:

• lock TTL: 60 seconds

If a lock exceeds its TTL, treat it as stale and reclaim it automatically before retrying the write. This prevents two models from waiting on abandoned locks forever.

---

Operational Principles

Parallelized Intelligence

Each cluster executes independently while synchronizing through shared state files.

Zero-Redundancy Coordination

No duplicated task execution across models.

Deterministic Routing

Optimus Prime dynamically routes workloads to the most specialized NIMs.

Security-by-Default

Red Alert validates every generation, mutation, and deployment path.

Persistent Project Memory

The 6-file architecture acts as the swarm’s distributed cognition layer.

---

Example Swarm Execution Flow

User Request
    ↓
Optimus Prime
    ↓
Task Classification
    ↓
Ultra Magnus → Architecture & Backend
Jazz → UI/UX
Ratchet → Testing & Refactoring
Red Alert → Security Validation
Perceptor → Retrieval & OCR
Bumblebee → Voice / Media
Ironhide → Simulation
Wheeljack → Scientific Reasoning
    ↓
6-File Synchronization Layer
    ↓
Unified Production Output