AI Daily Digest · 2026-08-29
🔬 New AI Craft
1. LM Studio breaks down Auto Review: AST parsing + capability extraction for safe agent shell execution
LM Studio's blog explains how Bionic's Auto Review mode parses shell commands with AST, extracts the "capabilities" a command invokes (file access, network, etc.), and matches them against policies before allowing execution — turning "what can this command do" into a statically analyzable, machine-checkable decision, unlike the manual-confirmation or prompt-constraint approach used in Codex / Claude Code.
https://lmstudio.ai/blog/how-auto-review-works
2. Anthropic researcher gives a peek at self-improving AI
TechCrunch covers an Anthropic researcher's preview of self-improving AI — models that iterate on their own training and capabilities. The recursive path of "models building better models" is far from production, but frontier labs are now openly discussing the engineering path.
https://techcrunch.com/2026/08/28/an-anthropic-researcher-just-gave-us-a-peek-at-self-improving-ai/
3. Autonomous mathematical discovery in an open-world multi-agent environment
A new arXiv paper proposes open-world multi-agent environments where agents collaboratively generate and prove math conjectures (75 points on HN) — a prominent example of agentic workflows applied to research automation, the same paradigm as agent-based task planning but aimed at math problems instead of product requirements.
https://arxiv.org/abs/2608.23691
🛠️ Tools & Tips
1. tokentab: cost-tracking CLI for Claude Code / Codex / Gemini CLI sessions
Reads agent session logs and tallies token spend by model, project, and day — handy for monitoring AI coding costs.
https://github.com/damejan80/tokentab
2. sepia: a "de-AI" writing skill for Claude Code and Codex
Narrative-architecture repair for fiction, making agent-written prose read like a human author.
https://github.com/Nanako0129/sepia
⭐ Open Source Highlights
1. rome-os/rome — "the agentic OS"
A new project positioning itself as an operating system for agents — orchestration, memory, and tool dispatch as OS-level abstractions — 384 stars in its first week.
https://github.com/rome-os/rome
2. Conduct — open-source guardrails for LLM and MCP tool calls
Filters and blocks unsafe or unauthorized tool invocations from LLM agents and MCP servers.
https://github.com/sseshachala/conductai
📰 Industry News
1. Federal judge rules Trump administration's blacklisting of Anthropic was illegal
2. China's daily AI token usage tops 500 trillion as compute demand grows
3. NSA official says the agency wants access to "all" AI models
🚀 Major Releases
1. Z.ai open-weights GLM-5.3; unsloth GGUF quantizations already available
2. Tencent unveils Hunyuan Hy4 Preview
