What is IntentusNet
IntentusNet is an open-source, language-agnostic execution runtime for multi-agent systems designed to make routing, fallback, and failure behavior deterministic, replayable, explainable, and production-operable.
The Core Premise
The model may change. The execution must not.
What IntentusNet Does
IntentusNet sits between your agent orchestration layer and the actual agent execution. It provides:
- Deterministic Routing — Given the same intent and set of agents, the router always selects agents in the same order
- Execution Recording — Every execution is captured as an immutable record with stable hashes
- Replay Without Re-execution — Recorded executions can be replayed to return the exact same output without calling models again
- Policy Filtering — Fine-grained control over which targets are allowed, with partial continuation support
- Structured Observability — Every routing decision produces trace spans and structured logs
What IntentusNet Is NOT
It's critical to understand what IntentusNet does not do:
| IntentusNet Does NOT | Use These Instead |
|---|---|
| Build agents | LangChain, AutoGen, CrewAI |
| Define prompts | Your agent framework |
| Call LLMs | OpenAI SDK, Anthropic SDK, etc. |
| Orchestrate complex workflows | Temporal, Prefect, or agent frameworks |
| Provide authentication | Your infrastructure layer |
| Load balance | Kubernetes, your cloud provider |
IntentusNet is a runtime layer that ensures deterministic execution semantics beneath whatever agent framework you choose.
The Runtime Analogy
Think of IntentusNet as systemd for AI agents:
- systemd doesn't write your service code
- systemd ensures reliable startup, restart, and supervision semantics
- systemd provides structured logging and dependency management
Similarly:
- IntentusNet doesn't write your agents
- IntentusNet ensures deterministic routing, crash recovery, and replay semantics
- IntentusNet provides execution recording and structured observability
When to Use IntentusNet
IntentusNet is designed for production multi-agent systems where you need:
- Reproducible debugging — "Why did the system choose that agent?" must have a deterministic answer
- Audit trails — Regulatory or compliance requirements for AI decision logging
- Failure recovery — Crash mid-execution shouldn't leave the system in undefined state
- Behavioral stability — Model upgrades shouldn't change routing decisions silently
When NOT to Use IntentusNet
IntentusNet may be overkill if:
- You have a single agent with no routing complexity
- You don't need execution replay or audit trails
- You're prototyping and don't care about production semantics yet
- Your system is stateless and doesn't need crash recovery
Version and Status
| Property | Value |
|---|---|
| Current Version | 1.3.0 |
| Python Support | 3.9+ |
| Status | Production-ready for core guarantees |
| License | MIT |
Core Dependencies
IntentusNet has minimal core dependencies:
cryptography >= 42.0.0
websockets >= 12.0
pyzmq >= 25.1.0
fastapi >= 0.110.0
uvicorn >= 0.29.0
python-dateutil >= 2.9.0
Quick Example
from intentusnet import IntentusRuntime, BaseAgent, IntentEnvelope, AgentResponse
# Create runtime with recording enabled
runtime = IntentusRuntime(enable_recording=True)
# Register agents (your agent implementation)
# ... agent registration code ...
# Execute intent - deterministically routed
response = runtime.router.route_intent(envelope)
# Response includes execution_id for replay
print(f"Execution ID: {response.metadata['execution_id']}")
print(f"Replayable: {response.metadata.get('replayable', False)}")
Next Steps
- Why Deterministic Execution Matters — Understand the problem IntentusNet solves
- MCP Compatibility — How IntentusNet relates to the Model Context Protocol
- Getting Started — Install and run your first intent