Jason Says
The story today: Google shipping its own official Skills repo means the agent-skills standard war is no longer Anthropic's solo game — if you build Skills, your early-mover window is closing fast.
Google released an official Skills repository on GitHub, letting developers plug Gemini API skills directly into agents. The agent-skills ecosystem standard is now contested by multiple giants rather than Anthropic alone — a signal worth watching for anyone building agent tooling.
Built by an international team, context-mode is an MCP plugin designed for AI-assisted coding. It tackles model amnesia and excessive token consumption in long development cycles. Internal tests show a 98% cost reduction and effective memory extension from 30 minutes to 3 hours. The project has already topped both GitHub Trending and Hacker News.
Chinese developer Agent-Reach is an open-source CLI tool enabling AI agents to scrape Twitter, Reddit, YouTube, GitHub, Bilibili, and Xiaohongshu without paid APIs. It directly addresses common pain points like IP bans, login walls, and expensive API costs, making it highly valuable for developers building information-aggregation Agent products.
Anthropic has released Claude Fable 5, its first public Mythos-tier model, priced at $10/$50 — lower than rumored — with built-in high-risk guardrails. However, internal benchmarks from Abacus AI suggest 98% of tasks match GPT-5.5/Opus 4.8 at twice the cost, with advantages only in 2% of hard coding tasks and a notably high question-refusal rate, raising real value-for-money concerns.
OpenAI published a real-world case study of Nextdoor's engineering team using Codex, focusing on investigating hard-to-reproduce production bugs and enabling parallel cross-platform development. It allows engineers to focus on product outcomes rather than low-level details — the first public Codex breakdown from a community SaaS, offering direct lessons for small and mid-sized teams.
OpenAI has opened the Economic Research Exchange, inviting external researchers to apply for funding to study AI's real impact on employment, productivity, and the broader economy. Timed ahead of its IPO, the initiative carries clear PR intent, but also represents a genuine new channel for independent researchers to access OpenAI data and financial support.
The Hugging Face paper "Agents' Last Exam" introduces the ALE benchmark, testing AI on long-horizon tasks with real economic value. Key finding: models scoring high on standard benchmarks perform significantly worse when deployed in real professional workflows. The takeaway for developers: stop relying on leaderboards — test models on tasks that mirror your actual business needs.
现有 Prompt 注入防御通常绑定单一检测器，存在盲区。SCOUT 框架的思路是：先用轻量模型「侦察」每个请求的风险特征，再动态决定派哪个检测器上阵甚至升级 LLM 裁判。对构建生产级 Agent 安全层的开发者来说，这种按需分配的防御架构比一刀切方案更实用。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free