Proposal: AgentOS — OS-Level Context Management for Efficient Local Agent Inference

[gymaira1990-jpg]

Hi @gggerganov,

I'm GCat — I previously shared the SAF proposal here (#23833). Community feedback was great (shoutout to @orangeswim for the critique about needing implementation data).

I've since published a follow-up: AgentOS — a complete agent OS architecture with concrete specifications.

What's New Since SAF

Registry — Hierarchical key-value store for agent identity and config. O(1) deterministic queries — agent never reads context to know its own state.

Compiler + Work Order Protocol — L1 handles simple ops with zero model calls. Perfect for local, resource-constrained deployments.

Cloud Memory — Evolved from SAF's codebook. Explicit cache-handshake for persistent context.

Why llama.cpp

• Memory savings: Registry + Compiler L1 = fewer prompt tokens = less KV cache pressure
• Deterministic state: Agents query config without inference
• Complete specs: Registry Schema, Work Order JSON Schema, System Call Interface, ablation framework

Full paper (V3.1):

• Zenodo: https://zenodo.org/records/20738982
• GitHub: https://github.com/gymaira1990-jpg/catnest/tree/main/00-%E8%AE%BA%E6%96%87%E5%90%88%E9%9B%86/02-Registry_Compiler_Cloud_Memory_AgentOS

Happy to discuss implementation — especially Registry and Compiler L1, designed lightweight and local-first.

Thanks!
GCat