AI Daily Digest · 2026-09-04
🔬 New AI Craft
1. Which tools do Claude Code, Codex and Cursor reach for? Armature measured 16,893 sessions to find out
The study quantifies how the three mainstream coding agents differ in when and which tools they invoke, treating tool selection itself as an observable, optimizable variable rather than a black box. Unlike the Plan-mode workflow you use daily, this is behavioral research on the agent side — the findings can directly inform prompt design and which tools you expose to steer an agent's trajectory.
https://armature.tech/blog/which-tools-coding-agents-install
2. Porting a 1993 Amiga game to Godot, with an LLM reading the 68000 assembly
A hands-on report on using an LLM as a "reverse-engineering assistant" — chunking 30-year-old Motorola 68000 assembly and feeding it to the model piece by piece, with human-in-the-loop control-flow checks — to port a legacy game engine to Godot. A concrete playbook for AI-assisted migration of ancient codebases, where the model does the bulk of mechanical translation and the human verifies control flow.
https://babyloniantwins.com/blog/porting-a-1993-amiga-game-to-godot/
3. ai-software-factory: GitHub issues in, merged PRs out — no human reads the diff
An open-source pipeline that turns issues into auto-merged pull requests behind layers of gates (tests, review), making machine-produced code defensible without human diff-reading. Quality here is enforced by automated checkpoints baked into the workflow, echoing the AI code-review philosophy rather than post-hoc human review.
https://github.com/coleam00/ai-software-factory
🛠️ Tools & Tips
1. reverify: an anti-hallucination layer for AI agents that read binaries
The model proposes, deterministic tools decide — every claim is VERIFIED or REFUTED against real bytes. Ships as an MCP server + CLI to stop agents confidently misreading binaries and disassembly.
https://github.com/2akouwu/reverify
⭐ Open Source Highlights
1. useagent: open-source AI coworker with its own cloud computer
Team-grade agents get a cloud machine, your tools and context, running Claude Code/Codex/OpenCode and handing back finished websites, decks, reports and PRs.
https://github.com/useagenthq/useagent
2. SlopMonster: lint AI slop out of copy, then cleanse it with a rival model
Detect AI tells, rewrite with a competitor model, lint again — for shipping copy that reads human.
https://github.com/ItsssssJack/SlopMonster
📰 Industry News
1. OpenAI, Claude and Grok went down simultaneously; SpaceXAI apologizes for the outage affecting Grok and other "compute partners"
2. Google Antigravity's TOS sparks backlash: third-party tool usage can get your Google account suspended
3. NYC mayor bans AI in schools through 8th grade for one year, citing foundational skills
🚀 Major Releases
1. OpenAI unveils GPT-6 Astra — 62.7% on ARC-AGI-3 with the standard harness, surpassing the human baseline
2. IFM open-sources K2 Horizon, a connected fleet of six models (375B flagship), Apache 2.0 across the whole training lifecycle
3. Qwen3.8 27B hits Cerebras at 1500 tokens/s
