Google Bets on Agentic Commerce, Anthropic Raises $65B, and Microsoft Brings AI Into Cloud Operations
Welcome to the latest edition of MCP Weekly, covering major developments from May 24th to May 31st, 2026. This week brought model launches, infrastructure releases, and the largest private funding round in AI history.
TL;DR
Anthropic closed a $65 billion Series H at a $965 billion valuation, with Claude now live across AWS, Google Cloud, and Azure. Claude Opus 4.8 launched with dynamic workflows that run hundreds of parallel subagents for large codebase migrations. Google shipped the Google Pay MCP Server and updated its payment platform for agentic commerce.
While open-sourcing Agent eXecutor (AX), a Kubernetes-native distributed runtime for long-running agents. Microsoft embedded SRE tools into the Azure MCP Server, letting developers manage cloud operations from their IDE. OpenAI published its Frontier Governance Framework and launched Rosalind Biodefense for pandemic preparedness.
Major Updates of the Week
Google Pay Goes Agentic
Google launched the Google Pay and Wallet Developer MCP Server, letting agents inside Cursor, VS Code, and other compatible tools access live merchant configurations, validate payment tokens, and manage integrations without switching context. Alongside this, Google updated its Pay platform at Google I/O to support the Universal Commerce Protocol, making existing merchant backends compatible with autonomous AI shopping agents without any rebuilding. Android now supports one-click Express checkout via dynamic callbacks. Google is wiring its payment infrastructure for an agent-first commerce world.
Claude Opus 4.8 Launches with Parallel Subagent Workflows
Claude Opus 4.8 ships with more than a benchmark update. It is four times less likely to let code flaws pass silently versus Opus 4.7, and introduces dynamic workflows in Claude Code that deploy hundreds of parallel subagents in a single session. In a real test, the system migrated 750,000 lines of code from Zig to Rust over eleven days with 99.8% of tests passing. Effort controls let users adjust reasoning depth per task, and the Messages API now accepts mid-session instructions without breaking running workflows. Project Glasswing’s Mythos model is in cybersecurity testing ahead of wider release.
Anthropic and Claude Updates
| Vendor / Product | Key Action / Feature | Significance |
| Anthropic Series H | $65B raised at a $965B valuation. Run-rate revenue at $47B. Multi-gigawatt compute deals with Amazon, Google, and SpaceX Colossus. | Claude is the first frontier model natively available across all three major clouds. |
| Anthropic Korea | KiYoung Choi appointed Representative Director ahead of a Seoul office. South Korea uses Claude at 3.5x the expected rate. SK Telecom and Law&Company already in production. | One of the fastest-growing Claude markets globally, now getting a dedicated local team. |
| Anthropic Milan | Sixth European office opened. JAKALA deployed Claude across 3,000 seats. Satispay compressed an 18-month roadmap to seven months. | Local presence ahead of EU AI Act enforcement across finance, pharma, and energy sectors. |
| Zero Trust for AI Agents | Published a three-tier Zero Trust framework for enterprise agents covering identity, sandboxing, task-scoped permissions, and agentic incident response. | AI now compresses the window from vulnerability discovery to exploit to hours. Standard perimeter security cannot keep pace. |
| LLM Source Code Security | Six-phase methodology for using Claude to find, verify, triage, and patch code vulnerabilities autonomously at scale. | Addresses the find-and-fix gap. Anthropic’s own repo had 1,596 open vulnerabilities with only 97 patched as of May 22nd. |
| CodeRabbit on Claude | Built a planning layer on Claude that produces a product requirements document before code generation, routing tasks across Opus, Sonnet, and Haiku by complexity. | Moves quality checks upstream. Cuts the costly rework that comes from AI coding in the wrong direction. |
Microsoft Puts SRE Agents Inside the Developer’s IDE
Microsoft added SRE Agent tools to the Azure MCP Server via the @azure/mcp package. Developers can now provision agents, manage incidents, configure connectors, and generate architecture plans directly from VS Code, Cursor, GitHub Copilot CLI, or Claude Desktop with no separate portal. Destructive actions require an explicit confirmation flag, tokens are scrubbed before entering any log, and access is gated by standard Azure role assignments. Cloud operations and software development now live in the same window.
OpenAI Updates
| Vendor / Product | Key Action / Feature | Significance |
| Rosalind Biodefense | Sponsored access program for GPT-Rosalind, a model built for biological research. Partners include Lawrence Livermore, Johns Hopkins, and CEPI’s 100 Days Mission for vaccine development. | First large-scale AI deployment aimed at pandemic preparedness and biosecurity defense at institutional level. |
| Frontier Governance Framework | Public governance document mapping internal safety protocols to California’s Transparency in Frontier AI Act and the EU AI Act. Covers cyber offense, CBRN risks, and incident response. | Shifts from voluntary ethics guidelines to an audit-ready compliance framework ahead of regulatory deadlines. |
| Tax AI with Thrive Holdings | Self-improving Codex agent for tax preparation across 30+ firms. Processed 7,000 returns, cut manual time by a third, raised field accuracy from 25% to 86% in six weeks by converting practitioner corrections into new evaluation targets. | A working example of an agent that improves itself in production from real human feedback. |
| Third-Party Evaluation Playbook | Framework for independent AI evaluations detailing how harness design and token budgets change measured results. Expanding a token budget from 10M to 100M tokens raised GPT-5.5 cyber eval scores by up to 59%. | The harness is as important as the model. Matters for anyone commissioning or reading AI safety evaluations. |
Google Open-Sources a Distributed Agent Runtime
Google released Agent eXecutor (AX) under Apache 2.0, a Go-based distributed runtime for long-running agents on Kubernetes. AX runs tools, skills, and agents as isolated actors with a single-writer event log that handles state recovery after interruptions automatically. It supports MCP and agent-to-agent communication natively, includes human-in-the-loop confirmation for high-risk actions, and lets developers fork and replay agent trajectories via CLI. This is production infrastructure for agents that run for hours or days, not a research demo.
Other Updates
| Vendor / Product | Key Action / Feature | Significance |
| AdRoll MCP Server | Beta MCP server lets agents run campaign reporting, pull intent data, and stage ad campaigns for human review. Works with Claude, ChatGPT, Cursor, and n8n. | Draft-first design keeps humans in control of spend while giving marketing teams conversational access to live data. |
| Base MCP | MCP server connecting Base crypto accounts to AI tools for token swaps and DeFi operations. Non-custodial. Agents stage transactions for user approval rather than executing directly. | Brings agentic workflows into DeFi without handing the AI access to private keys. |
| Higher Logic Vanilla MCP | MCP server giving AI tools access to live community sentiment, product feedback, and support trends. Free for existing customers. Respects existing roles and audit trails. | B2B teams can query customer community data from Claude or ChatGPT without manual exports. |
| Detectify MCP Server | Remote MCP server feeding live vulnerability scan data and attack surface findings into AI coding tools. Agents can ingest findings as structured remediation tasks. | Connects external security scanning to the same tools writing the code, closing the gap between finding and fixing. |
| Red Hat Kiali MCP | Kubernetes MCP server in Technology Preview integrating Kiali mesh telemetry into AI tools. Supports live topology views, trace analysis, and conversation-driven Istio fixes via OpenShift Lightspeed. | Cluster operators can troubleshoot service mesh issues through a chat interface instead of parsing dashboards during an outage. |
My Thoughts: The Compute Story
The $65 billion Anthropic raise is easy to view as another funding headline. It is more revealing as a compute story. Multi-gigawatt agreements across AWS, Google Cloud, and Azure, alongside memory supply commitments from Micron, Samsung, and SK hynix, suggest Anthropic is planning for a future where agent workloads consume infrastructure at a fundamentally different scale than today’s chat applications.
The more interesting pattern is what the industry is building around those models. Google’s Agent eXecutor, Microsoft’s Azure SRE agents, Anthropic’s Zero Trust framework, and Google’s agentic commerce initiatives are all solving different pieces of the same problem: how to run agents that persist, recover from failures, operate within governance boundaries, and interact with real-world systems safely.
A year ago, most agent announcements focused on what models could do. This week’s releases focused on what happens after deployment. Runtime environments, operational controls, approval workflows, recovery mechanisms, and infrastructure for long-running execution are increasingly becoming the areas where vendors are differentiating.
The frontier models are converging. The supporting infrastructure is not. That may prove to be the more important competitive battleground over the next few years.
