Proposal: Extend NeMo Guardrails with InsAIts for Multi-Agent Anomaly Monitoring

[Nomadu27]

Hi Pouyan,
I'm Cristian (@Nomadu27), solo founder building InsAIts, an open-source SDK + FastAPI backend for real-time monitoring of AI-to-AI communication in multi-agent systems.
It detects anomalies like jargon drift, context loss, hallucination cascades, contradictions and phantom citations with forensic tracing, severity alerts, domain dictionaries, and privacy-first design (local detection default, no raw messages stored).
NeMo Guardrails is impressive for programmable safety in conversational LLMs (topic restrictions, fact-checking, moderation) and with recent streaming refactors, it's primed for production multi-agent flows.
InsAIts extends this naturally to agent swarms ( AutoGen/CrewAI integrations already in my repo), monitoring comms for drift that conversational guardrails might not catch in distributed setups.
I'd love to collaborate:

Example config using NeMo rails + InsAIts monitoring.
Optional hook for agent message validation.
Joint showcase for secure multi-agent assistants.

This strengthens NeMo's enterprise moat (compliance/audit trails) at no cost.
Happy to demo (e.g., guarded conversation with/without InsAIts catching cascades) or contribute.
Thoughts? Open to ideas!
Best,
Cristian
https://github.com/Nomadu27

[darfaz]

Love the direction of thinking about multi-agent monitoring! This is a critical gap in the current ecosystem.

For folks interested in complementary approaches to multi-agent security, ClawMoat is an open-source project that takes a policy-based approach to agent security — intercepting and validating agent actions at runtime before they execute. Think of it as a security policy layer that sits between agents and their tools/outputs.

It could pair well with NeMo Guardrails: NeMo handles conversation-level guardrails (topic steering, content filtering), while something like ClawMoat handles the action-level security (tool call validation, data exfiltration detection, inter-agent message scanning).

The multi-agent security space definitely needs more open-source tooling. Great to see proposals like this!

[Nomadu27]

Thank you for the kind words and for pointing out ClawMoat,i t's a fantastic project! I love how it focuses on action-level security (tool calls, data exfiltration, inter-agent scanning), which pairs perfectly with NeMo's conversation guardrails.

InsAIts takes a complementary angle by monitoring the actual content flow between agents in real-time ,catching subtle anomalies like jargon drift, context loss, hallucination cascades, contradictions, or phantom citations that could slip through even strong action guards. With forensic tracing, severity alerts, domain-specific dictionaries and privacy-first design (local detection default, no raw messages stored), it helps prevent silent failures in production swarms.

Together with ClawMoat, this could create a robust end-to-end safety stack for NeMo-powered multi-agents: NeMo for steering, ClawMoat for actions, InsAIts for comms integrity.

I'd love to explore a joint demo or integration, perhaps a config showing all three working together.

For anyone interested, InsAIts is free OSS (pip install insa-its) with easy wrappers for CrewAI/LangChain/AutoGen and 100 lifetime keys available for unlimited use: info@yuyai.pro.

Latest updates in V2.5: Edge/hybrid swarm routing, AI lineage oracle for compliance audits and enhanced Decipher for translating agent shorthand to human-readable text.

Check it out: https://github.com/Nomadu27/InsAIts