Summary

The day's signals converge on one shift: AI agents are moving from answering questions to autonomously carrying out multi-step work. Meta launched Muse, a consumer personal agent that books, buys, and pays with your card across email, web, and WhatsApp, pushing payment-capable autonomy into mass distribution. On the developer side, Cursor's new Projects mode fields a persistent coordinator that plans work and delegates to parallel cloud subagents, while Vercel added GitHub Copilot to its AI SDK harness layer so apps can hot-swap rival coding agents behind one interface. Google extended Gemini to a native Windows desktop app with an Alt + Space overlay, and OpenHands 1.0 shipped as the top open-source coding scaffold with production-grade sandboxing. Together they mark agents becoming autonomous, orchestrated, and ubiquitous across consumer and developer surfaces.

Key themes

  • From answering to acting: agents now execute real-world tasks and transactions autonomously, most starkly with Meta Muse holding a user's card to book, negotiate, and pay across web, apps, and WhatsApp.
  • Orchestration over single assistants: platforms are converging on managed coordination of many coding agents — Cursor Projects with a delegating coordinator and parallel cloud subagents, and Vercel's harness layer making agent choice a config detail rather than a rewrite.
  • Provider-neutral interoperability: Vercel's @ai-sdk/harness-github-copilot puts Copilot, Claude Code, Codex, Cursor, and others behind one HarnessAgent interface over the Agent Client Protocol, reducing lock-in.
  • Assistants at the OS level: Google's Gemini Windows app puts an AI overlay one keystroke away with direct Workspace context, contesting Microsoft Copilot on its home turf.
  • Open scaffolds close the gap: OpenHands 1.0 pairs Docker sandboxing and security rails with leaderboard-topping, model-agnostic SWE-bench results, sharpening the build-vs-buy calculus against paid coding agents.

Notable items

  • Meta Muse — consumer personal AI agent (US) that reads email, books travel, fills forms, negotiates, and pays with the user's card via muse.ai, iOS/Android, and WhatsApp; free with Power ($20/mo) and Maximum ($100/mo) tiers. High impact; sets a consumer pricing anchor as agents move from answering to acting.
  • Cursor Projects — a persistent coordinator agent that plans and delegates coding work to parallel cloud subagents, keeps context over long horizons, runs on its own cloud machine, and performs recurring work unprompted. High impact for engineering leaders.
  • Vercel AI SDK harness for GitHub Copilot — @ai-sdk/harness-github-copilot connects Copilot through the same HarnessAgent interface (via ACP), letting apps swap coding agents (Claude Code, Codex, Cursor, Cline, Grok Build, OpenCode) without code changes.
  • Google Gemini desktop app for Windows 10/11 — native x64/ARM64 client with an Alt + Space overlay, Workspace (Gmail/Drive) context, and on-desktop Nano Banana image generation; follows the earlier macOS release.
  • OpenHands 1.0 — open-source coding scaffold with production-ready Docker sandboxing, built-in security policies, resource limits, and a plugin system; model-agnostic SWE-bench Verified from 46.8% (Devstral 24B) to ~72% (Claude Sonnet 4.5, extended thinking), the top open scaffold, with 40k+ GitHub stars.

Source coverage

Source rows used: 5