Cloud vendors adopt session-first agent runtimes

AWS, Microsoft, Google, and Anthropic have converged on session-based execution for AI agent runtimes, launching nearly identical managed architectures within six weeks in spring 2026. The vendors agree on the session as the primary unit but differ on isolation methods, with AWS using hardware-level microVM sandboxing and Microsoft and Google using software-level sandboxing. A portable AGENTS.md and SKILL.md config format, backed by the Linux Foundation, is emerging as a standard across platforms.

According to The New Stack, AWS, Microsoft, Google, and Anthropic have converged on the session - not the request - as the primary unit of execution in managed AI agent runtimes. Three vendors shipped nearly identical managed runtime architectures within six weeks in spring 2026: Anthropic launched Claude Managed Agents on April 8 at $0.08 per session hour , AWS extended Bedrock AgentCore on April 22, and Google followed with Managed Agents via Antigravity at I/O. Each collapses model API, sandbox, orchestration, and hosting into configuration files, with per-session isolation. The vendors now agree the session is the right unit but disagree on how to isolate it - AWS uses hardware-level microVM sandboxing, while Microsoft Foundry and Google use software-level sandboxing. The AGENTS.md and SKILL.md config format is emerging as a portable standard across all three platforms, stearded by the Linux Foundation and present in over 60,000 open-source repositories.