Today's through-line is AI moving from answering questions to doing work — and being asked to explain itself while doing so. Anthropic opened a public forum on the thorniest societal questions about AI, OpenAI upgraded ChatGPT's agent mode with persistent memory and native developer tool integrations, and the GitHub trending leaderboard was dominated entirely by agentic skills frameworks aimed at making coding agents more capable in production.
Anthropic Launches "Inviting Hard Questions" Initiative
Anthropic published a new initiative explicitly asking the public to surface the toughest questions about AI — things like "who decides the rules?" and "does AI make the world more dangerous?" — and committed to showing their reasoning as they work through each one. It's a notable departure from the typical AI company posture of speaking only when ready. Whether or not it shifts public opinion, having a frontier lab publicly commit to an ongoing process of deliberation is a meaningful signal of how much scrutiny the industry knows it's operating under.
anthropic.com — Inviting hard questions
Claude Gets a Reflection Dashboard
Anthropic also launched a beta reflection feature in Claude's Settings — available on the web and desktop apps — that lets users visualize how they use Claude over time: what kinds of tasks they bring to it, how patterns shift, where they spend their time. Most AI interfaces are black boxes from the user's side, and giving people a mirror on their own usage opens up questions about habits and dependencies that most AI products have kept deliberately opaque. It's a small move, but an interesting one.
anthropic.com — A new way to reflect on how you use Claude
ChatGPT Agent Mode v2: Persistent Memory and Dev Tool Connectors
OpenAI shipped a significant upgrade to ChatGPT's agent mode: persistent memory across sessions, the ability to schedule recurring tasks, and native connectors for GitHub and Jira rolling out to Plus and Team subscribers. The GitHub and Jira integrations are the most consequential piece for developers — they move the agent from a separate tool you context-switch to, toward something embedded in the actual development workflow. The question agents now have to answer is how much of a development loop they can own end-to-end, not just assist with at the margins.
mattpocock/skills: Engineering Skills Built for Agents
The top trending GitHub repo today at +1,712 stars, mattpocock/skills packages engineering skills for use directly from a .claude directory — structured, reusable, and aimed at real-world coding tasks rather than toy examples. As agentic coding tools mature, the bottleneck is shifting from raw model capability to the quality of the task definitions agents work from. This project is an attempt to set a higher bar for what "skill" means in a production engineering context, and it's clearly hitting a nerve.
obra/superpowers: An Agentic Development Methodology
Close behind with +1,013 stars today, obra/superpowers takes a broader approach than a pure skill library — it frames itself as both a framework and a software development methodology for agentic work. The combination has resonated: it sits somewhere between a tools repo and a playbook for how teams should structure work when agents are doing meaningful parts of the coding. Alongside mattpocock/skills, it's part of a visible trend of developers investing seriously in the scaffolding that makes agentic workflows repeatable and predictable rather than one-off magic tricks.