intentusnet 1.3.0

IntentusNet is a deterministic AI intent routing and execution runtime that makes LLM-driven systems inspectable, replayable, and failure-safe.

IntentusNet

Deterministic Execution Runtime for Intent Routing and Multi-Agent Systems

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

IntentusNet is an open-source, language-agnostic execution runtime for multi-agent systems.
It makes routing, fallback, and failure handling deterministic, replayable, explainable, and production-operable.

IntentusNet focuses strictly on execution semantics, not planning or intelligence — ensuring that execution behavior remains predictable even when models are not.

---

---

Why IntentusNet

Modern LLM systems are observable, but not debuggable.

In real production systems, failures are often:

• irreproducible
• incorrectly blamed on models
• hidden behind retries and fallback logic
• impossible to replay or audit

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

• Replayable
• Attributable
• Explainable

---

Execution Recording & Deterministic Replay

IntentusNet treats executions as immutable facts, not transient logs.

Each execution is:

• recorded as a first-class artifact
• replayable deterministically (without re-running models)
• inspectable after crashes or upgrades

This enables:

• reliable root-cause analysis
• auditability and compliance
• safe model iteration without rewriting history

The model may change.
The execution must not.

This design 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, reason about goals, or optimize prompts.

---

Core Capabilities

• Deterministic intent routing
• Explicit fallback chains
• Execution recording (WAL-backed)
• Deterministic replay & verification
• Crash-safe recovery
• Typed failures & execution contracts
• Operator-grade CLI
• Transport-agnostic execution

---

Intent-Oriented Routing

• Capability-driven routing
• Explicit fallback sequences
• Sequential or parallel execution
• Priority-based routing
• Auditable routing decisions
• Trace spans with execution metadata

Routing decisions are deterministic and replayable, not heuristic.

---

EMCL Secure Envelope (Optional)

IntentusNet supports EMCL (Encrypted Model Context Layer):

• AES-GCM authenticated encryption
• HMAC-SHA256 signing (demo provider)
• Identity-chain propagation
• Anti-replay protections

EMCL is optional and transport-agnostic.

---

MCP Compatibility

IntentusNet is MCP-compatible by design:

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

IntentusNet provides deterministic execution semantics around MCP tools.

---

Language-Agnostic Design

Agents can be implemented in any language that supports:

• HTTP / JSON
• ZeroMQ
• WebSocket

Including: Python, C#, Go, TypeScript, Rust

---

SDK Status

Included — Python Runtime SDK

• Intent router & fallback engine
• Agent base classes
• Agent registry
• Multi-transport execution
• Execution recorder & replay engine
• WAL-backed crash recovery
• EMCL providers
• MCP adapter
• CLI tooling
• Example agents & demos

Note:
Higher-level ergonomic SDKs (decorators, auto-registration) and C#/TypeScript SDKs are planned next.

---

Demos

deterministic_routing_demo

Compares three approaches using identical capabilities:

• without — ad-hoc production glue code
• with — deterministic routing via IntentusNet
• mcp — routing backed by a mock MCP tool server

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

execution_replay_example

Shows how model upgrades change live behavior while past executions remain replayable.

---

Operational Scope (Important)

IntentusNet is a deterministic execution runtime, not an autonomous agent system.

Guarantees

• Deterministic routing, fallback, and failures
• Crash-safe execution recording
• Deterministic replay or loud failure on divergence
• Explicit contracts and typed failures
• CLI-first operational control

Explicit Non-Goals

• No task planning or reasoning
• No evaluation of model outputs
• No replacement for workflow engines
• No distributed consensus in v1

Determinism Boundary

Determinism is enforced at the execution layer, not the model layer.
Non-deterministic model behavior is detected, recorded, and surfaced — never hidden.

---

Roadmap

Next

• Python ergonomic SDK
• C# SDK
• TypeScript SDK
• MCP adapter improvements
• EMCL key rotation

Future (Optional)

• Multi-agent planning layer (research)
• Trust-scored routing

---

Author

Balachandar Manikandan

---

License

MIT License — open-source

---

Keywords

Deterministic execution runtime, intent routing, explicit fallback chains, replayable agent workflows, debuggable LLM systems, execution recording, WAL-backed recovery, MCP-compatible runtime, EMCL-secured agent communication, transport-agnostic AI infrastructure.