AI Daily Digest · 2026-06-23
🔬 New AI Craft
1. Sakana Fugu — Multi-Agent System as a Model
Sakana AI unveils Fugu, a model interface that coordinates specialized AI agents through a single OpenAI-compatible API. Unlike your daily Agent Plan workflow — Fugu treats the entire multi-agent system as one "model" behind a unified API, so you can switch between single-model and multi-agent setups without changing application code.
https://sakana.ai/fugu/
2. You're Probably Using Agent Skills Wrong
Anson Biggs breaks down proper Agent Skill patterns and common antipatterns — directly relevant to how you structure skills for agent workflows, covering skill granularity, context management, and cross-skill dependencies.
https://notes.ansonbiggs.com/youre-probably-using-agent-skills-wrong/
3. Prompt Injection as Role Confusion
New research demonstrates prompt injections work by exploiting the model's role assignment mechanism, not just by overriding instructions. Provides a new conceptual framework for understanding AI security.
https://role-confusion.github.io
🛠️ Tools & Tips
1. Oak — Version Control Designed for AI Agents
A Git alternative purpose-built for agents, with branching and merge semantics optimized for agent workflows instead of human collaboration.
https://oak.space/oak/oak
2. Unsloth Adds GLM-5.2 Local Inference Support
Run GLM-5.2 efficiently on consumer hardware with gguf quantization and Unsloth optimizations.
https://unsloth.ai/docs/models/glm-5.2
3. Codex Logging Bug Can Destroy Your SSD
Codex SQLite feedback logs can write up to 640 TB/year, rapidly consuming SSD endurance. OpenAI merged a fix reducing logs by 85%.
https://github.com/openai/codex/issues/28224
4. Claude Code's Extended Thinking Output Isn't "Authentic"
Analysis argues Extended Thinking text is a simplified post-hoc narrative, not the model's actual reasoning process — similar to the thinking chains you see in Agent Plan, it's presentation, not cognition.
https://patrickmccanna.net/the-text-in-claude-codes-extended-thinking-output-is-not-authentic/
⭐ Open Source Highlights
1. omnigent-ai/omnigent (4.4K⭐)
Open-source meta-harness orchestrating Claude Code, Codex, Cursor, and custom agents from a single interface.
https://github.com/omnigent-ai/omnigent
2. apple/coreai-models (1.1K⭐)
Apple's open-source on-device AI model toolchain with Python primitives and Swift runtime for edge inference.
https://github.com/apple/coreai-models
3. Forward-Future/loop-library (1K⭐)
Library of reusable AI-agent loop patterns, installable as a Skill, covering common agent workflow templates.
https://github.com/Forward-Future/loop-library
📰 Industry News
1. Geohot warns "The Doom Justifies the Valuation" — AI valuation bubble analysis
2. Nvidia launches Halos, a full-stack AV safety system from cloud to car
3. Chevron signs 20-year power agreement with Microsoft for Texas AI data center
🚀 Major Releases
1. GLM-5.2 vs Opus benchmark: GLM-5.2 matches Claude Opus on coding tasks
2. Moebius: 0.2B parameter image inpainting framework delivers 10B-class performance
3. Deno Desktop launches, bringing Deno runtime to desktop GUI applications
