The last post here (After Google Next, the Agentic Era Got Specific) ended with one line: MCP and A2A are the substrate. Ten days later, the two biggest labs in the world made that line load-bearing — in opposite ways.
Google used I/O 2026 to publish a coherent open stack for agents: a fast model built for the loop, an agent-first IDE, a managed-agent runtime, a proposed web standard, and a consumer agent stitched across all of it.
Anthropic spent the same week building walls — acquiring the SDK generator every rival used and shutting it down, lining up a ~$900B funding round, and extending Opus 4.7 into security.
Two strategies, one prize. Here's what each is actually buying.
| Google (Open) | Anthropic (Moat) | |
|---|---|---|
| The Bet | Publish the full stack, win on developer mindshare | Acquire the choke points, lock in structural advantage |
| Signature Move | WebMCP proposal + Antigravity 2.0 IDE + Spark | $300M+ Stainless acquisition, shut down for rivals |
| Distribution | Workspace + Spark in Gmail, Docs, Slides | Vertical platforms (finance, legal, security) |
| Why It's Rational | Needs developer mindshare to replace search-ad revenue | $900B valuation without ad/consumer-distribution channel |
Google's Open-Stack Bet
The I/O keynote shipped roughly 100 announcements. Filter to the agent stack and a coherent platform emerges, layer by layer.
Model layer — Gemini 3.5 Flash. Frontier intelligence at Flash speeds. It outperforms Gemini 3.1 Pro on coding and agentic benchmarks while running roughly four times faster than other frontier models. This is the first major release explicitly positioned as the model for agentic loops, where round-trip latency matters more than the last percentage point on a reasoning benchmark.
IDE layer — Antigravity 2.0. The agent-first development environment, rebuilt around subagents, hooks, asynchronous task management, and multi-agent orchestration as primitives. Google is treating "writing agents" as a distinct discipline from "writing code that happens to call an LLM" — and shipping tooling that assumes that distinction.
Runtime layer — Managed Agents in the Gemini API. One API call spins up an agent that reasons, uses tools, and executes code in an isolated Linux environment. This is the same primitive Anthropic shipped a few weeks ago as Managed Agents on Claude Platform. Two of the three frontier labs now offer the same building block under nearly the same name. That's convergence at the runtime layer.
Standard layer — WebMCP. A proposed open standard for exposing JavaScript functions and HTML forms as agent-callable tools. The important word is proposed. WebMCP is not adoption; it is a public bet that the browser becomes part of the agent runtime. Worth tracking. Not yet worth depending on. But Google publishing a standard, in public, is the strategic posture.
Consumer layer — Gemini Spark. Covered in its own section below.
The shape of the bet: Google is doing for agents what they did for Android. Make the full stack available, mostly open, and win by being the easiest place to build. Win developer mindshare, win distribution.
Anthropic's Moat Bet
In the same week Google was publishing standards, Anthropic was acquiring infrastructure.
On May 18, Anthropic acquired Stainless — the SDK-generation startup founded by ex-Stripe engineer Alex Rattray — for a reported $300M+. Stainless takes API specifications and produces production-ready SDKs across Python, TypeScript, Kotlin, Go, and Java, then keeps them updated as the underlying APIs change. It is the tool OpenAI, Google, Cloudflare, Replicate, and Runway all used to ship and maintain their developer experience. Anthropic bought it and is winding down the hosted product. Existing customers keep the SDKs they've already generated. Nobody else gets to use the tool going forward.
The right way to read this is not "Anthropic acquired a dev-tools startup." It is "Anthropic identified the integration-layer choke point and bought it." The integration layer is where agents connect to external software — exactly the surface Anthropic, OpenAI, and Google are all racing to standardize. Removing a shared tool from the shared pool is a moat move, not a product move.
Pair this with Claude Security, which Anthropic moved into public beta the same week. Code vulnerability scanning and proposed fixes built on Opus 4.7, scheduled and targeted scans, triage tracking. It extends the vertical-by-vertical strategy from the last post — finance, legal, security, all wrapped in Anthropic platform plus pre-built agents plus deep data partnerships.
The valuation context — a ~$900B funding round closing the week of May 26, which would surpass OpenAI's $852B — makes the spend rational. At that price, structural lock-in matters more than feature velocity. The Stainless deal is the cheapest moat money you can buy when one of your rivals just published a proposed open standard.
Spark and the Distribution End
Gemini Spark is the consumer end of Google's I/O stack — and it is the announcement most worth taking seriously.
Spark is a 24/7 personal agent built on Gemini 3.5 and the Antigravity harness, integrated into Gmail, Docs, Slides, and Workspace generally, with launch partners including Canva, OpenTable, and Instacart. Because it runs as a cloud agent, it keeps working after you close the laptop. You teach it tasks — flag deadlines in Gmail, summarize long threads, spot hidden fees in monthly credit card statements — and it executes them on its own schedule. Available to Google AI Ultra subscribers starting the week of May 25.
Two reads matter for builders.
First, distribution is the moat Google already has. Anthropic can buy Stainless. It cannot buy Workspace. Spark works because Google can ship the stack end-to-end into accounts where the user already is. That is the advantage no funding round can replicate.
Second, "personal agent" stopped meaning chat. Spark works in the background, on a schedule, against your real data. That is the bar the rest of the industry now has to meet. "Personal agent" as a synonym for "chatbot with memory" is no longer enough. Expect Apple, Microsoft, and Notion to ship counters in the next quarter.
The Money Behind Both Bets
Both strategies become rational when you look at the valuations underwriting them.
Anthropic's $30B-plus funding round is expected to close the week of May 26 at a pre-money valuation above $900B — surpassing OpenAI's $852B March valuation. Q1 revenue reported at $4.8B, with annualized run rate projected past $50B by the end of June. Roughly an 80x year-over-year growth and a 15x valuation jump in fourteen months.
OpenAI filed its IPO prospectus confidentially with the SEC on May 22, targeting a September listing.
At those numbers, the strategic logic of each lab becomes legible.
Google's open-stack bet is rational because Google needs developer mindshare on a scale large enough to justify replacing search-ad revenue with agent-economy revenue. Open standards and free-tier tooling are how you get that mindshare.
Anthropic's moat bet is rational because Anthropic does not have ads or a consumer distribution channel of comparable scale. Without those, a $900B price tag has to be justified by structural lock-in — vertical platforms, exclusive data partnerships, and integration-layer ownership.
Different exits force different strategies. The bets are not contradictions. They are the rational moves at each lab's valuation.
What This Means for Builders
Three things to do with this map.
Inventory your toolchain. Stainless going dark is a preview. If a tool in your agent stack is venture-backed and AI-lab-adjacent, it can be acquired and shut down on a week's notice. Know what depends on what. Pin SDK versions. Be ready to fork or replace.
Track WebMCP. Don't bet on it yet. A proposed standard from Google is worth watching, but it is eighteen-plus months from real adoption even on the optimistic path. MCP and A2A still win on adoption today. Build to those. Treat WebMCP as a future option, not a present plan.
Distribution is the next frontier. Spark is what happens when an agent stack meets pre-existing distribution. The lab with the best stack does not automatically win. The lab that ships the stack into accounts users already use — Workspace, Microsoft 365, Notion — wins. Watch for Microsoft's equivalent on top of Agent 365, and Notion's on top of the External Agent API. The next big competitive shifts happen there, not at the model layer.
The agent stack is now built. The fight that matters is over where it lives.
Sources: I/O 2026 developer highlights (Google), 100 things we announced at Google I/O 2026, Google introduces Gemini Spark (TechCrunch), Anthropic acquires Stainless, Anthropic Funding Round to Top $30B (Tech Times)