AI Daily Digest · 2026-08-16
🔬 New AI Craft
1. Codex auto-research drives a 232x kernel speedup
The author let Codex take over the full kernel optimization loop in "auto-research" mode — locating bottlenecks, forming hypotheses, writing and running experiments, with a human only making the final call — landing a 232x performance gain. Unlike the Agent Plan mode you use, the agent here runs the complete research-hypothesis-experiment cycle instead of executing pre-split tasks; the human shifts from writing code to reviewing research conclusions.
https://sankalp.bearblog.dev/autoresearch/
2. AI-assisted GPU porting of a 250k-line legacy weather simulation code
A new arXiv paper demonstrates an AI-assisted workflow for porting 250k lines of legacy code to GPU, offering a reproducible recipe for modernizing legacy systems — the kind of "old code to GPU" job where agents outlast humans.
https://arxiv.org/abs/2608.13122
3. Working with AI feels more like leadership than coding
Hot HN post (256pt): effective AI collaboration demands goal-setting, task decomposition, and outcome review — leadership, not coding skill, which matches the "break down requirements, review acceptance" rhythm of your Agent Plan workflow.
https://allen.bargi.org/notes/working-with-ai-feels-like-leadership/
4. agent-safe-pipeline: agents propose, but cannot authorize
Open-source reference architecture where agents can propose actions but cannot authorize their own execution — immutable intent capture plus a separate approval component hard-codes permission boundaries into the pipeline, built for agent automation that needs strong auditing.
https://github.com/decionis/agent-safe-pipeline
🛠️ Tools & Tips
1. eve Software Factory template (Vercel)
Vercel's official template to spin up an eve software factory with Foreman in one go.
https://github.com/vercel-labs/eve-software-factory-template
2. airship: a Figma-like visual editor for Claude Code/Codex
Visually choreograph agent workflows for Claude Code, Codex, and OpenCode, Figma-style.
https://github.com/0xnyn/airship
3. tokentab: know what your coding agents really cost
A CLI that reads Claude Code, Codex, and Gemini CLI session logs and breaks down spend by model and project.
https://github.com/wzchav/tokentab
⭐ Open Source Highlights
1. watermarks-remover: 9.7k stars in 5 days for stripping AI watermarks
Removes multi-vendor AI provenance marks (C2PA metadata, Unicode text fingerprints) — the most contested trending project right now.
https://github.com/guillaumemeyer/watermarks-remover
2. MasterAgent: agents that run 100% on-device
Sub-100ms latency on Qualcomm NPU, zero cloud dependency.
https://github.com/OpenSparX/MasterAgent
3. moli: an agent-first browser written in Rust
A lightweight Rust browser designed for AI agent interaction and automation.
https://github.com/lexmount/moli
📰 Industry News
1. Alibaba's AI models pass 3B downloads, overtaking Meta and Google
🚀 Major Releases
1. Anthropic shares more details on how Claude's new watermarks will work
