SMESH

Decentralized LLM coordination · Plant-inspired · Blazing fast

Quick Start • What Is This? • Performance • Comparison

---

🚀 Quick Start

# Build
cargo build --release

# Check everything works
cargo run --bin smesh -- status

# Run a simulation (watch signals flow!)
cargo run --bin smesh -- sim --nodes 50 --ticks 100

# Compare LLM backends (Ollama vs Claude)
ANTHROPIC_API_KEY=... cargo run --bin smesh -- compare

---

🤔 What Is This?

TL;DR: Coordinate multiple LLM agents without a central controller.

Traditional (LangChain)	SMESH
Central orchestrator routes everything	No center — agents self-organize
One failure breaks the chain	Fault tolerant — signals just decay
10-50ms per coordination step	~1μs per signal operation

How It Works

Think plant communication, not message queues:

   Agent A emits          Signals decay           Agent B senses
   "need review"          over time               & claims task
        ↓                     ↓                        ↓
      ████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░████████
      
   Multiple agents agree → Signal gets reinforced → Consensus emerges

Key ideas:

• 📡 Signals diffuse through a shared field
• ⏱️ Signals decay — stale tasks fade away
• 🤝 Reinforcement — agreement strengthens signals
• 🌿 Emergent — no explicit voting needed

---

⚡ Performance

What	Speed
Signal creation	216 ns (4.6M/sec)
Signal reinforcement	48 ns (20M/sec)
100-node network tick	14 μs

LLM Backends

Backend	Latency	Cost
Ollama (local)	0.1-1s	Free
Claude (API)	1-4s	Pay per token

---

🏗️ Architecture

smesh-rust/
├── smesh-core/      # Signals, nodes, fields, networks     ✓
├── smesh-runtime/   # QUIC P2P networking                  ✓
├── smesh-agent/     # Ollama + Claude backends             ✓
└── smesh-cli/       # CLI tools & benchmarks               ✓

All components complete. 80 tests passing.

---

⚔️ vs Other Tools

vs LangChain

	LangChain	SMESH
Speed	~10-50ms/step	~1μs/signal
Architecture	Central DAG	Decentralized
Failure mode	Chain breaks	Signals decay
Memory	High (Python)	Low (Rust)

vs CrewAI / AutoGPT

	CrewAI	SMESH
Scaling	Linear	Sublinear
Consensus	Explicit voting	Emergent
Control	Role prompts	Trust + decay

---

✅ When to Use

Good for:

• 3+ LLM agents coordinating
• Fault-tolerant systems
• High-throughput (many signals/sec)
• Edge deployment (single binary)

Not for:

• Single chatbot
• Strict sequential workflows
• Exactly-once guarantees

---

📚 Core Concepts

Concept	One-liner
Signal	Message with intensity that decays over time
Node	Entity that emits, senses, and reinforces signals
Field	Shared space where signals live and propagate
Network	Topology connecting nodes (ring, mesh, small-world, etc.)

---

🔧 CLI Commands

smesh status              # Check system & Ollama connection
smesh sim                 # Run signal simulation
smesh agents --demo       # LLM agent coordination demo
smesh compare             # Benchmark Ollama vs Claude
smesh bench               # Signal processing benchmarks
smesh review --path ./repo    # SMESH-coordinated code review
smesh code                # Multi-agent coding swarm (Claude)
smesh swarm --path ./repo # Vulnerability scanning swarm

---

🔍 Code Review (NEW)

Multi-agent code review using signal diffusion:

cargo run --bin smesh -- review --path ./some-repo --model qwen2.5-coder:7b

Agent	Focus
Security	Unsafe code, vulnerabilities, input validation
Performance	Allocations, algorithms, hot paths
Style	Idiomatic Rust, patterns, organization
Documentation	Doc comments, examples, clarity

Agents emit findings as signals → related reviewers reinforce → consensus emerges.

---

🌿 Coding Swarm (NEW)

Multi-agent collaborative coding that demonstrates true SMESH coordination:

cargo run --bin smesh -- code
cargo run --bin smesh -- code --coders 3 --format json

Agent	Role
Architect	Designs module structure, defines interfaces
Coder (2x)	Implements code from specs
Tester	Writes and runs tests
Reviewer	Reviews code quality

SMESH concepts in action:

• 📡 Signal coordination — Agents emit Task, Claim, Code, Review, TestResult signals
• 🤝 Emergent consensus — Modules complete when review + tests agree
• 📊 Trust evolution — Trust scores change based on code acceptance and test results

Trust Evolution:
  Reviewer→Coder-A [███████   ] 0.70  (code approved)
  Tester→Coder-A   [████      ] 0.45  (1 pass, 1 fail)
  Tester→Coder-B   [█         ] 0.20  (2 fails)

---

📄 License

MIT OR Apache-2.0 — pick whichever works for you.