- OpenAI added “Record & Replay” to Codex on macOS. A user can walk Codex through a workflow once, such as uploading a YouTube video with metadata, thumbnail, and subtitles, and Codex turns it into a reusable skill it can run again. The update also adds bulk actions for automation history and thread handoff between local and remote hosts.
- ChatGPT’s scheduled tasks are becoming more like a personal assistant layer. A new “Scheduled” page collects active tasks in one place, where users can view, pause, edit, or delete them. Research tasks can check the web and connected apps, then alert users only when something changes.
- Data2Story turns a CSV into a verifiable interactive news article through a seven-agent workflow. The Oxford and Stanford system splits the newsroom pipeline into roles such as Detective, Analyst, Editor, Designer, Programmer, Auditor, and Inspector. Its most interesting feature is the Inspector panel, which links claims, charts, and interactive elements back to code, data, or external URLs.
- OpenAI reportedly brought in $5.7 billion in Q1 2026 revenue while burning $3.7 billion. Revenue and cash burn both tripled year over year, and operating loss reached $9.3 billion according to the report. The story highlights the central AI business tension: demand is growing, but compute costs and price competition remain hard constraints.
- John Jumper, the Nobel-winning AlphaFold scientist, is leaving Google DeepMind for Anthropic. Jumper spent nearly nine years at DeepMind and helped lead the AlphaFold work that won the 2024 Nobel Prize in chemistry. TechCrunch also notes he had been involved in Google’s coding tools efforts.
- Hugging Face’s MosaicLeaks study shows how deep-research agents can leak private information through search queries. When agents combine local private documents with external web search, harmless-looking queries can add up into a “mosaic” of sensitive facts. A simple “don’t leak” prompt was not enough; the proposed PA-DR privacy-aware RL method reduced leakage while improving strict chain success.
- Hugging Face also published a benchmark for testing whether tools are truly agent-friendly. The benchmark does not just ask whether an agent got the right answer; it measures how many turns, tokens, seconds, and detours it needed. The core lesson is practical: clear CLIs, structured examples, and discoverable docs can make software much cheaper and safer for agents to operate.
Bottom line: the agent era is moving from demos to operations, where repeatability, verifiability, privacy, and cost all have to be engineered together.