What happened
The New Stack reports that AWS, Microsoft, Google, and Anthropic have independently converged on session-aware managed agent runtimes, with the session replacing the request as the primary execution unit. A related New Stack analysis (May 27, 2026) documented that Anthropic, AWS, and Google shipped nearly identical managed runtime shapes within six weeks: Anthropic launched Claude Managed Agents in public beta on April 8 (priced at $0.08 per session hour), AWS extended Bedrock AgentCore on April 22, and Google introduced Managed Agents via Antigravity at Google I/O in late May. Microsoft's Foundry Agent Service, expected to reach general availability in early July 2026, adds a fourth session-first runtime to the picture.
The convergence
Per The New Stack, all four vendors now ship managed runtimes that collapse model API, sandbox, orchestration, and hosting into configuration and a small number of API calls. Each one spins up an isolated environment per agent session - AWS AgentCore uses dedicated microVM sandboxes with isolated CPU, memory, and filesystem; Microsoft Foundry runs each session in its own sandboxed environment; Google Antigravity spins a remote Linux sandbox per agent invocation; Anthropic scopes credentials, memory, and tool access to the session.
The disagreement
Where the vendors diverge is on isolation mechanism. AWS favors hardware-level microVM isolation (stronger security guarantees, higher latency floor). Google and Microsoft use software-level sandboxing (faster cold starts, different security tradeoffs). The practical effect for engineering teams: the isolation choice determines both the latency budget for session startup and the trust model for multi-tenant deployments.
Emerging config standard
Per The New Stack, AGENTS.md and SKILL.md are emerging as the portable configuration format across platforms. The AGENTS.md format is stewarded by the Linux Foundation and appears in over 60,000 open-source repositories. Google Antigravity, Anthropic Managed Agents, and AWS AgentCore all support variants of this file-based config. The practical effect: an agent definition can be diffed and checked into Git with no proprietary DSL.
What it means for practitioners
Per The New Stack, the runtime is now table stakes - the vendor decision moves to data location, session-hour pricing, model choice, and ease of migration. Industry-pattern observations: session-first execution eliminates context resets between turns but introduces new operational concerns - session affinity routing, per-session cost accounting, and multi-tenant isolation guarantees become first-class infrastructure requirements. Teams building production agents should now evaluate runtimes on those "boring" criteria rather than on runtime availability alone.
What to watch
Isolation standard maturation - whether hardware-level isolation (microVM) becomes the default for regulated industries, or software sandboxing closes the gap on security guarantees. Config portability - whether AGENTS.md/SKILL.md remains genuinely cross-platform or fragments into vendor-specific dialects. Pricing dynamics - Anthropic's $0.08/session-hour benchmark sets a reference price; watch whether Google and AWS match it or differentiate on compute tiers.
Key Points
- 1AWS, Microsoft, Google, and Anthropic all now offer session-first managed agent runtimes, making the runtime itself table stakes rather than a vendor differentiator.
- 2Vendors disagree on isolation: AWS uses hardware-level microVM sandboxing while Google and Microsoft use software-level sandboxing, creating different latency and security tradeoffs.
- 3AGENTS.md and SKILL.md are emerging as a portable, Linux Foundation-stewarded config format across platforms - the decision for practitioners now shifts to pricing, data location, and migration ease.
Scoring Rationale
A New Stack synthesis documenting four major cloud vendors converging on session-first managed agent runtimes and their disagreement on isolation mechanisms. Solid infrastructure analysis relevant to practitioners building production agents, but the underlying product launches (Anthropic April 8, AWS April 22, Google May 27) occurred weeks prior - this is pattern-recognition reporting rather than a new launch.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
