Summary

May 16 centered on AI systems becoming more operationally embedded and more governable at the same time. The strongest signals combined wider agent distribution into enterprise workflows with tighter controls around execution, routing, and security: Cursor pushed agents into Microsoft Teams and hardened cloud environments, Vercel added policy-style controls across routing and protected access, Anthropic paired enterprise expansion with a sharper safety claim, and OpenAI extended both Codex supervision and ChatGPT’s product surface.

Key themes

  • Agent workflows kept moving closer to where enterprise teams already work, with Cursor in Microsoft Teams, Codex on mobile, and Databricks positioning GPT-5.5 inside existing orchestration layers.
  • Governance and security were a major throughline, with Cursor adding environment-level controls, Vercel locking down source maps by default, and Anthropic publishing a more concrete agent-safety benchmark.
  • Platform operators gained more runtime control rather than just model access, including Vercel routing by cost, latency, or throughput and authenticated CLI testing for protected deployments.
  • Enterprise distribution stayed central, highlighted by PwC scaling Claude Code and Cowork across a much larger services footprint.

Notable items

  • Cursor launched a Microsoft Teams integration for delegating work to cloud agents directly from team threads.
  • Cursor added version history, rollback controls, audit logs, and scoped secrets for cloud-agent development environments.
  • Vercel shipped Protected Source Maps, plus new AI Gateway sorting controls and authenticated vercel curl access for protected deployments.
  • Anthropic said Claude models since Haiku 4.5 scored 0% on its agentic misalignment blackmail eval, while PwC expanded Claude Code and Cowork deployment at firm scale.
  • OpenAI expanded Codex with mobile remote access and enterprise access tokens, and separately previewed Plaid-linked personal finance workflows inside ChatGPT.

Source coverage

Source rows used: 10