Physics-based coding agent. Minimizing action in Ω-space.
H = T + V (Hamiltonian = kinetic + potential)
S = ∫(T−V)dt → min (variational principle)
The world's first coding agent built on SWML (Shunsuke's World Model Logic) — a mathematical framework that describes AI agent behavior using Hamiltonian mechanics and variational principles.
Tasks are energy states. The agent finds the path of least action — just like nature.
███████╗██╗ ██╗███╗ ███╗██╗
██╔════╝██║ ██║████╗ ████║██║
███████╗██║ █╗ ██║██╔████╔██║██║
╚════██║██║███╗██║██║╚██╔╝██║██║
███████║╚███╔███╔╝██║ ╚═╝ ██║███████╗
╚══════╝ ╚══╝╚══╝ ╚═╝ ╚═╝╚══════╝
A G E N T
Every other coding agent is a glorified while-loop: prompt → tool call → repeat.
SWML-Agent operates in Ω-space — a state space where each phase has a measurable Hamiltonian:
| Phase | T (kinetic) | V (potential) | H (total) | Meaning |
|---|---|---|---|---|
| 🔭 OBSERVE | 0.3 | 0.9 | 1.2 | Low action, high uncertainty |
| 📐 PLAN | 0.5 | 0.5 | 1.0 | Building structure, reducing entropy |
| ⚡ EXECUTE | 0.9 | 0.2 | 1.1 | High action, low uncertainty |
| ✓ VERIFY | 0.2 | 0.1 | 0.3 | Ground state. Convergence. |
The agent minimizes total action S = ∫(T−V)dt — the same principle that governs photon paths, planetary orbits, and quantum field theory.
╔════════════════════════════════════════════════════════╗
║ 🔭 OBSERVE → 📐 PLAN → ⚡ EXECUTE → ✓ VERIFY ║
║ ║
║ H=1.10 T=0.9 ▓▓▓▓▓▓▓▓▓▓▓▓▓░░ V=0.2 ▓▓▓░░░░░░░░░ ║
║ S=-0.342 steps=3 tools=5 t=14.2s ║
║ η=78% entropy=1.58 tokens=1240→890 ║
║ task: Create a Python web server with health check ║
╚════════════════════════════════════════════════════════╝
Plus an ASCII energy trajectory plot showing H(t) over time.
# 1. Install Ollama
# https://ollama.com/download
# 2. Pull a model
ollama pull qwen3-coder # 30B — best quality
# or
ollama pull qwen3:8b # 8B — faster, lighter
# 3. Run
python swml-agent.pyNo pip install. No node_modules. No API keys. One file.
# Interactive mode
python swml-agent.py
# One-shot
python swml-agent.py -p "Create a REST API with SQLite backend"
# Specify model
python swml-agent.py --model qwen3:8b
# Auto-approve all tool calls (yolo mode)
python swml-agent.py -y
# Switch model mid-session
Ω ❯ /model qwen3:8b- 🔬 Ω-space state tracking — Hamiltonian decomposition (H = T + V) in real-time
- 📐 Action integral — S = ∫(T−V)dt measures computational "effort"
- 📊 Efficiency metric — η = how close to optimal path (%)
- 📈 Energy trajectory — ASCII plot of H(t) over time
- 🔀 Variational planner — Scores multiple plans by predicted action
- 🛠️ 10 built-in tools — read, write, edit, patch, run, list, search, think, checkpoint, undo
- 💾 Git checkpoints — Auto-save before risky changes, undo to rollback
- 🧠 Think tool — Internal reasoning without side effects
- 📦 Zero dependencies — Python stdlib only, single file (~900 lines)
- 🏠 Fully local — Ollama backend, no cloud, no cost
In classical mechanics, nature finds the path that minimizes action:
S = ∫ L(q, q̇, t) dt where L = T - V
T = kinetic energy (how hard the agent is working) V = potential energy (how much uncertainty remains) H = T + V (total energy, the Hamiltonian) S = total action (the integral we minimize) η = efficiency (1 - |S_actual| / |S_worst|)
A good agent:
- Spends minimal time in high-V states (confusion, exploration)
- Applies high-T actions only when V is already low (directed execution)
- Converges quickly to ground state (VERIFY, H=0.3)
This is the variational principle — the deepest optimization algorithm in physics, now applied to software engineering.
| Command | Description |
|---|---|
/status |
Show current Ω-state and trajectory |
/models |
List available Ollama models |
/model X |
Switch to model X |
/undo |
Rollback last git checkpoint |
/help |
Show help |
/quit |
Exit |
| Setup | Model | Speed | Quality |
|---|---|---|---|
| 8GB VRAM | qwen3:8b | ~50 tok/s | ★★★ |
| 12GB VRAM | qwen3-coder (30B Q4) | ~10-15 tok/s | ★★★★★ |
| 16GB+ VRAM | qwen3-coder (30B Q8) | ~20-30 tok/s | ★★★★★ |
| 24GB VRAM | qwen3-coder (30B FP16) | ~30-40 tok/s | ★★★★★ |
Works on Mac (Apple Silicon), Windows (NVIDIA), and Linux.
| Claude Code | vibe-local | SWML-Agent | |
|---|---|---|---|
| Cost | $20+/mo + API | Free | Free |
| Local | ✗ | ✓ | ✓ |
| Dependencies | Node.js | Python | Python (stdlib only) |
| Lines of code | ~100k+ | ~7400 | ~900 |
| Theoretical basis | None | None | SWML (Hamiltonian mechanics) |
| Energy tracking | ✗ | ✗ | ✓ (H, S, η, entropy) |
| Git checkpoints | ✗ | ✓ | ✓ |
| Variational planning | ✗ | ✗ | ✓ |
- vibe-local by Yoichi Ochiai — proved single-file local agents work
- SWML Theory — the math behind this agent
- Hamilton, Lagrange, Feynman — the variational principle
Shunsuke Hayashi (@swml_lab)
Physics M.Sc. → Factory optimization → Amazon operations management → AI agent architecture.
The same math that describes particle physics describes AI agents. This is the proof.
MIT