Secure Runtime for Intent Routing and Multi-Agent Execution

Deterministic • Transport-Agnostic • EMCL-Ready • MCP-Compatible

IntentusNet is an open-source, language-agnostic execution runtime for multi-agent systems, designed to make routing, fallback, and failures deterministic, debuggable, and production-operable.

It focuses on execution semantics — not planning or intelligence — ensuring that routing, fallback, and failure behavior remain predictable even when models are not.

---

🎯 Why IntentusNet Exists

Modern LLM systems are observable, but not debuggable.

Failures are often:

• irreproducible
• incorrectly attributed to models
• impossible to replay
• obscured by implicit retries and hidden fallback logic

IntentusNet addresses this by enforcing deterministic execution semantics around LLMs, allowing failures to become:

• Replayable
• Attributable
• Explainable

This design philosophy is formalized in:

📄 RFC-0001 — Debuggable Execution Semantics for LLM Systems
→ rfcs/RFC-0001-debuggable-llm-execution.md

Non-goals: IntentusNet does not plan tasks, generate tools, or control model reasoning.

---

⚠️ Python SDK Notice:
The Python Runtime SDK (router, agents, transports, EMCL, MCP adapter) is included in this release.
A higher-level ergonomic SDK (decorators, auto-registration, PyPI package) arrives in the next version.
C# SDK also arrives next version.

---

✨ Key Features

• Language-agnostic agent model
• IntentEnvelope abstraction for clarity
• Agent registry + capability schema
• Deterministic routing with fallback support
• Identity-aware agent execution
• EMCL envelope encryption (optional)

Supported Transports:

• HTTP / JSON
• ZeroMQ
• WebSocket
• In-process
• (Future) MCP-native transport

---

⚡ Intent-Oriented Routing Engine

• Capability-driven routing
• Envelope-defined fallback chain
• Sequential or parallel intent flows
• Priority-based routing
• Trace spans with metadata
• RouterDecision for auditing workflows

---

🔐 EMCL Secure Envelope (Optional)

• AES-GCM authenticated encryption
• HMAC-SHA256 signing (demo provider)
• Identity-chain propagation
• Nonce/timestamp/anti-replay logic

---

🔗 MCP-Compatible Architecture

The architecture is compatible with MCP:

• Agents can be wrapped as MCP tools
• Accept MCP tool requests
• Emit MCP-style responses
• Optional EMCL-secured MCP envelopes

The MCP Adapter is included in the runtime.

---

🌐 Language-Agnostic Design

Agents can be written in:

• Python
• C#
• Go
• TypeScript
• Rust
• Any language speaking HTTP/ZeroMQ/WebSocket

---

📦 SDK Status

✔️ Included in This Release — Python Runtime SDK

• Agent base class
• Router + fallback engine
• AgentRegistry
• Transports: ZeroMQ, HTTP, WebSocket, In-process
• EMCL providers (AES-GCM, HMAC)
• MCP Adapter
• Protocol models & validators
• Trace sink
• Example agents & demos

---

🧪 Demos (What Changes in Practice)

IntentusNet demos focus on code structure and execution behavior, not AI output quality.

The primary demo is:

deterministic_routing_demo

This demo compares three real-world approaches using the same logical capabilities:

• without — typical production glue code
  Manual tool calls, explicit retries, hand-written fallback logic.

• with — centralized routing using IntentusNet
  Routing, fallback, and traceability are expressed as routing options and handled by the router.

• mcp — same routing flow backed by a mock MCP tool server
  Demonstrates local + remote agents participating in the same routing model.

The demo is intentionally non-interactive by default and safe to run in Docker or CI.

python -m examples.deterministic_routing_demo.demo --mode without
python -m examples.deterministic_routing_demo.demo --mode with
python -m examples.deterministic_routing_demo.demo --mode mcp

---

📌 Coming Next Version

Python Ergonomic SDK

• @agent decorators
• Auto-capability registration
• Schema validation helpers
• PyPI packaging
• Workflow utilities

C# SDK

• Routing client
• Agent interfaces
• EMCL provider
• Transports

---

### 🧰 Architecture Snapshot

┌──────────────────────────────────────────────────────────────┐
│                    Client / Application                      │
└──────────────────────────────────────────────────────────────┘
                               │
                          Intent Router
                               │
                       Agent Execution Layer
                               │
                  EMCL Secure Envelope (Optional)
                               │
┌──────────────────────────────────────────────────────────────┐
│ MCP Tools │ ZeroMQ │ HTTP │ WebSocket │ In-Process           │
└──────────────────────────────────────────────────────────────┘
           Backend Agents / Services (Any Language)

📦 Capabilities Summary

Included

• Intent router + fallback
• Capability matching
• Multi-transport execution
• EMCL encryption
• Trace spans
• Agent identity + correlation IDs
• MCP-compatible core
• Python runtime SDK

Planned

• Python ergonomic SDK
• C# SDK, TypeScript SDK
• Full MCP Adapter (inbound + outbound)
• EMCL key vault + rotation

---

🗺 Roadmap

Next Release

• Python ergonomic SDK
• C# SDK
• TypeScript SDK
• MCP Adapter
• EMCL key rotation

Future

• Multi-agent planner (research / explicitly optional)
• Trust-scored routing

---

👤 Author

Balachandar Manikandan

---

📍 License

MIT License — Open source & commercial friendly.

---

🔐 Keywords

AI agent execution runtime, deterministic execution semantics, intent routing, explicit fallback chains, debuggable LLM systems, agent routing engine, execution traceability, replayable agent flows, MCP-compatible runtime, MCP tool execution, Model Context Protocol, distributed agent execution, transport-agnostic agent runtime, secure agent communication, EMCL encrypted envelopes, ZeroMQ agent transport, WebSocket agent transport, production AI infrastructure, LLM orchestration runtime