The agent layer is no longer a research project. It is a product category with real revenue, real protocols, and real distribution deals.

Google has spent the first quarter of 2026 shipping an interlocking stack of protocols and platforms: discovery (Universal Commerce Protocol), inter-agent communication (Agent2Agent), orchestration (Vertex AI Agent Builder), payments (Agent Payments Protocol), and enterprise deployment (Gemini Enterprise, formerly Agentspace). It is building the full infrastructure, top to bottom.

Anthropic has taken the opposite approach. It shipped the most capable autonomous model (Claude Opus 4.6), developer tools (Claude Code, now at $2.5 billion in annualised revenue), a consumer agent product (Cowork), and an open integration standard (Model Context Protocol). Then it embedded Claude inside Microsoft's M365 Copilot, reaching hundreds of millions of enterprise seats without owning a single one.

We mapped this competition when GPT-5.4 launched. The Google-Anthropic version of this race is structurally different. It is not a benchmark war. It is an architecture war.

Google's agent strategy is vertically integrated. Each layer connects to the others, and the whole stack runs on Google Cloud.

Universal Commerce Protocol (UCP) is the discovery layer. Announced at NRF in January 2026, UCP standardises how AI agents find products, compare options, and initiate purchases. It is already live in the US, with shoppers buying from Etsy and Wayfair through Google's AI Mode in Search and the Gemini app. Over 20 global partners endorsed UCP at launch, including Shopify, Target, Walmart, Adyen, Mastercard, Visa, and Stripe.Agent2Agent (A2A) is the communication layer. Now at version 0.3 and donated to the Linux Foundation, A2A enables agents from different vendors to negotiate with each other. Over 150 organisations support it. Google also launched an AI Agent Marketplace on Google Cloud, where partners can sell A2A-compatible agents directly to enterprise customers.

Vertex AI Agent Builder is the orchestration layer. It lets enterprises build, deploy, and govern multi-agent systems using the Agent Development Kit. Production-ready agents in under 100 lines of Python. HIPAA-compliant workloads. Customer-managed encryption. Private VPC deployment.

Agent Payments Protocol (AP2) is the settlement layer. Developed with more than 60 organisations including American Express, PayPal, Coinbase, and Worldpay, AP2 uses cryptographically signed "Mandates" as tamper-proof proof of user intent. It supports credit cards, debit cards, stablecoins, and real-time bank transfers.

Gemini Enterprise (formerly Agentspace) is the enterprise deployment layer. It bundles search, AI assistants, and agent creation into a single platform with prebuilt connectors for Confluence, Jira, SharePoint, and ServiceNow. A no-code Agent Designer lets non-technical employees build custom agents.

Anthropic's strategy looks nothing like Google's. There is no commerce protocol. No payment rails. No marketplace. Instead, Anthropic is building the best autonomous model and distributing it through everyone else's infrastructure.

Claude Opus 4.6 is the foundation. Released in March 2026, it holds the METR record for autonomous task completion: a 50 percent time horizon of 6.6 hours, meaning it can complete tasks that take a human roughly 6.6 hours about half the time without intervention. In one test, it autonomously closed 13 issues and assigned 12 to the right teams across six repositories in a single day. This is not a chatbot. It is a worker.

Claude Code is the developer product. At $2.5 billion in annualised revenue as of February 2026, it is Anthropic's fastest-growing product line, having more than doubled since the start of the year.

Claude Cowork is the consumer and enterprise agent. Launched on macOS in January and Windows in February, Cowork runs inside an isolated virtual machine, handling multi-step workflows across files, documents, and applications. Anthropic rolled out private plugin marketplaces and enterprise connectors for Google Drive, Gmail, DocuSign, FactSet, and more.

Model Context Protocol (MCP) is the integration standard. Now hosted by the Linux Foundation's Agentic AI Foundation, MCP has over 10,000 public servers registered and 97 million monthly SDK downloads. It is supported by OpenAI, Google, Microsoft, and Amazon. MCP is to agent tool integration what USB was to peripherals: a universal plug.

Distribution through partnerships is the go-to-market. On March 9, Microsoft launched Copilot Cowork, powered by Anthropic's Claude, inside Microsoft 365. The deal sits on top of a $30 billion Azure compute agreement. Claude is now available in Copilot Studio, in mainline Copilot chat, and as the agent engine behind multi-step workflows in Word, Excel, and PowerPoint. Anthropic is also available on AWS through Amazon Bedrock.

Both companies are competing for the same enterprise budgets, and their strategies collide in two areas.

Enterprise agent deployment. Google's Vertex AI Agent Builder and Anthropic's Claude (via Microsoft Copilot and AWS Bedrock) are both targeting the same procurement decision: which platform do we use to deploy agents across the organisation? Google offers a vertically integrated stack with governance built in. Anthropic offers model quality and the flexibility to run inside whatever stack the enterprise already uses.

Developer ecosystems. Google's A2A and Anthropic's MCP are both open standards, both donated to the Linux Foundation, and both aiming to become the default protocol for agent interoperability. A2A handles agent-to-agent communication. MCP handles agent-to-tool integration. They are complementary on paper. In practice, the company whose protocol becomes the default integration point gains enormous leverage over the agent ecosystem.

As we noted in our analysis of the agentic commerce stack, protocol wars tend to be winner-take-most. The standard that developers adopt first tends to stick.

The structural difference between these strategies is clear when you follow the money.

Google monetises infrastructure. Every agent built on Vertex, every transaction processed through AP2, every commerce interaction routed through UCP generates cloud revenue. Google does not need to build the best model. It needs to build the platform that the best models run on, including its own Gemini and, through A2A, agents powered by competitors.

Anthropic monetises intelligence. Every enterprise that embeds Claude, whether through Microsoft, AWS, or Anthropic's own API, pays for model usage. Anthropic's $14 billion revenue run rate and $380 billion valuation are built on being the model that enterprises trust for their hardest tasks. It does not need to own the platform. It needs to be the model every platform wants.

This creates a genuine strategic tension. Google's stack is powerful but locked to Google Cloud. Anthropic's model is portable but dependent on distribution partners who have their own AI ambitions. Microsoft is simultaneously Anthropic's biggest distribution partner and the company most likely to build a competing model.

The risk for Google is that the best agents do not run on its infrastructure. The risk for Anthropic is that its distribution partners eventually replace it with in-house models.

This is where the competition matters most for anyone in payments, commerce, or fintech.

Google has built the full commerce stack: UCP for discovery, AP2 for payments, A2A for agent negotiation. When an AI agent on Google's stack finds a product, negotiates a price, and initiates a payment, every step runs through Google's protocols. The Mandates system in AP2, with cryptographic proof of user intent, is Google's answer to the authentication and liability questions that we identified as unresolved in agent commerce.

Anthropic has built none of this. Claude can operate software, manage files, and execute multi-step workflows. But it has no commerce protocol, no payment rails, and no transaction authentication framework. When a Claude-powered agent needs to make a purchase, it will route through someone else's payment infrastructure.

That "someone else" is the critical variable. If Claude powers the agent inside Microsoft Copilot, the transaction likely routes through Microsoft's commerce infrastructure. If Claude powers an agent on AWS, it routes through whatever payment stack the enterprise has deployed. Anthropic's model touches the transaction but does not control it.

For Visa, Mastercard, and the major processors, Google's approach is a potential threat. A vertically integrated agent commerce stack that includes its own payment protocol could disintermediate existing rails. Anthropic's approach is more compatible with the current payments infrastructure, because it does not try to replace it.

The $650 billion question we have been tracking, who captures the value when AI restructures commerce, has a new dimension. It is no longer just about which company builds the best model. It is about which architecture controls the transaction layer.

AP2 adoption by payment networks. Google's Agent Payments Protocol has 60+ partners, but no live consumer transactions yet. Watch for Visa or Mastercard to either adopt AP2 or publish competing frameworks. The network that moves first on agent payment authentication will set the standard.

MCP vs A2A convergence. Both protocols are at the Linux Foundation. Both are open. The question is whether they stay complementary or whether one absorbs the other's use cases. If MCP expands into agent-to-agent communication, or A2A expands into tool integration, the protocol war becomes direct.

Anthropic's commerce play. Anthropic has no commerce protocol today. If agent-initiated transactions become a significant revenue category, that gap becomes a strategic liability. Watch for partnerships with payment networks or processors.

Microsoft's model loyalty. Microsoft is shipping Copilot Cowork with Anthropic's Claude and simultaneously investing in OpenAI. Enterprise customers will soon be choosing between Claude and GPT models inside the same Microsoft interface. Which model wins the enterprise default will determine where billions in agent-initiated transactions route.

Enterprise lock-in patterns. Google's vertically integrated stack creates switching costs. Anthropic's model portability reduces them. Watch which pattern enterprises prefer: the convenience of one stack or the flexibility of best-of-breed.

When agents control the transaction, does it matter more to own the pipes or to be the brain deciding what flows through them?