Skip to main content

Introducing IntentusNet - Deterministic Execution for Multi-Agent AI

· 2 min read
Balachandar Manikandan
Balachandar Manikandan
Creator of IntentusNet

Today we're releasing IntentusNet, a deterministic execution runtime for multi-agent AI systems. IntentusNet makes routing, fallback, and failure behavior replayable, explainable, and production-operable.

The Problem​

Multi-agent AI systems are fundamentally non-deterministic. When you deploy agents in production, you face:

  • Non-deterministic routing: Agent selection varies between runs
  • Silent failures: Errors disappear without trace
  • Unreplayable executions: Cannot reproduce issues
  • Debugging nightmares: "Why did the system do that?"

These aren't edge cases—they're daily realities for teams running AI agents in production.

Our Approach​

IntentusNet is a runtime layer that sits beneath agent frameworks to make execution deterministic and debuggable. It's not an agent framework—it doesn't define prompts, call LLMs, or orchestrate workflows. Instead, it ensures:

  1. Deterministic Routing: Same input → same agent selection, always
  2. Execution Recording: Every execution captured with stable hashes
  3. Replay Without Re-execution: Return recorded outputs, no model calls
  4. Policy Filtering: Fine-grained control with partial continuation

Quick Example​

from intentusnet import IntentusRuntime

# Create runtime with recording enabled
runtime = IntentusRuntime(enable_recording=True)

# Register your agents
# ... agent registration ...

# Execute intent - deterministically routed
response = runtime.router.route_intent(envelope)

# Later: replay returns exact same output
replay_result = runtime.replay(execution_id)

What's in v1.3.0​

  • Core Runtime: Deterministic routing, execution recording, replay engine
  • Policy Engine: Allow/deny rules with partial filtering
  • Multiple Transports: In-process, HTTP, WebSocket, ZeroMQ
  • MCP Adapter: Bridge to Model Context Protocol
  • Comprehensive CLI: Inspect, replay, validate, estimate

Philosophy​

IntentusNet is built on a core belief:

The model may change. The execution must not.

AI systems are inherently unpredictable. Models update, confidence scores shift, and outputs vary. But the execution infrastructure should be rock-solid—deterministic, auditable, and recoverable.

Think of it like systemd for AI agents: it doesn't write your service, but it ensures reliable execution semantics.

Get Started​

pip install intentusnet

Check out the documentation to get started.

What's Next​

We're working on:

  • WAL-backed persistence for stronger durability guarantees
  • Async routing support
  • Additional execution stores (PostgreSQL, Redis)
  • Enhanced cost estimation

Follow us on GitHub to stay updated.

IntentusNet is open source under the MIT license. We welcome contributions, feedback, and production war stories.