Llms Ai Agents

I cut Claude Code's token usage by 68.5% by giving agents their own OS

Reddit - Artificial Intelligence March 28, 2026 1 min read

About this article

Al agents are running on infrastructure built for humans. Every state check runs 9 shell commands. Every cold start re-discovers context from scratch. It's wasteful by design. An agentic JSON-native OS fixes it. Benchmarks across 5 real scenarios: Semantic search vs grep + cat: 91% fewer tokens Agent pickup vs cold log parsing: 83% fewer tokens State polling vs shell commands: 57% fewer tokens Overall: 68.5% reduction Benchmark is fully reproducible: python3 tools/ bench_compare.py Plugs into...

You've been blocked by network security.To continue, log in to your Reddit account or use your developer tokenIf you think you've been blocked by mistake, file a ticket below and we'll look into it.Log in File a ticket

Originally published on March 28, 2026. Curated by AI News.

Read Original Article

Llms

Claude is the least bullshit-y AI

Just found this “bullshit benchmark,” and sort of shocked by the divergence of Anthropic’s models from other major models (ChatGPT and Ge...

Reddit - Artificial Intelligence · 1 min · about 2 hours ago

Llms

HALO - Hierarchical Autonomous Learning Organism

The idea is called HALO - Hierarchical Autonomous Learning Organism. The core premise is simple: what if instead of just making LLMs bigg...

Reddit - Artificial Intelligence · 1 min · about 4 hours ago

Llms

[P] TurboQuant for weights: near‑optimal 4‑bit LLM quantization with lossless 8‑bit residual – 3.2× memory savings

An adaptation of the recent TurboQuant algorithm (Zandieh et al., 2025) from KV‑cache quantization to model weight compression. It gives ...

Reddit - Machine Learning · 1 min · about 5 hours ago

Llms

[D] Litellm supply chain attack and what it means for api key management

If you missed it, litellm versions 1.82.7 and 1.82.8 on pypi got compromised. malicious .pth file that runs on every python process start...

I cut Claude Code's token usage by 68.5% by giving agents their own OS

About this article

Related Articles

Claude is the least bullshit-y AI

HALO - Hierarchical Autonomous Learning Organism

[P] TurboQuant for weights: near‑optimal 4‑bit LLM quantization with lossless 8‑bit residual – 3.2× memory savings

[D] Litellm supply chain attack and what it means for api key management

No comments

Stay updated with AI News