Hello HN,

I’m Siva. I spent the last few months building Kernx, a high-performance Java runtime designed for deterministic AI agent workloads.

The Problem: Most "Agent Frameworks" are just Python wrappers around API calls. They suffer from unpredictable latency (GC pauses), context switching overhead, and "queue bloat."

The Solution: Kernx is a specialized kernel built on Java 25 (Preview) that treats compute as a deterministic pipeline. - Architecture: Single-process, multi-tenant. - Concurrency: 100% Virtual Threads (Project Loom). No reactive callbacks. - Memory: Foreign Function & Memory API (Panama) to bypass the Java Heap for data buffers. Zero GC pressure on the hot path.

The Results: On a standard local machine (MacBook Air M1), it sustains ~66,000 requests/second with sub-1ms p99 latency. It doesn't orchestrate containers; it just executes logic.

It is currently v1.0. I am looking for feedback on the architecture and the decision to use Scoped Values for memory safety.

Repo: https://github.com/Kernx-io/kernx Docs: https://kernx.io

Roast my code.