Google just declared the assistant era over. At I/O 2026, the company unveiled Gemini 3.5, the Antigravity 2.0 agent SDK, and WebMCP — an open web standard for browser-based AI agents. The message was unambiguous: the next computing platform isn't an app, a browser, or even a voice assistant. It's an autonomous agent.
Gemini 3.5 and the Antigravity 2.0 SDK
The headline model at this year's I/O was Gemini 3.5, which Google claims achieves state-of-the-art results on multi-step reasoning benchmarks while reducing inference latency by 40% compared to Gemini 2.0. But the more consequential announcement was the Antigravity 2.0 SDK — a CLI-first framework for orchestrating specialized subagents into production workflows.
Antigravity 2.0 ships with a new agent compose command that lets you declare a DAG of subagents in YAML, wire their inputs and outputs, and deploy the entire pipeline with a single command. Each subagent can be powered by a different model — you might route cheap classification tasks to Gemini Flash and complex reasoning to Gemini 3.5 Ultra — and the SDK handles retries, fallbacks, and observability automatically.
For developers who've been hand-rolling agent orchestration in LangChain or building custom loops with raw API calls, Antigravity 2.0 is worth a serious look. Google also open-sourced an Android Skills library that encodes Android development best practices into pre-built agent capabilities, letting LLMs handle common mobile dev tasks without needing to rediscover conventions from scratch.
WebMCP: Bringing Agents into the Browser
The announcement that generated the most developer chatter was WebMCP — a proposed open standard (currently at draft stage) that would allow browser-based AI agents to interact directly with JavaScript functions and HTML form elements on any webpage.
Think of it as a typed, permissioned API layer that sits between an AI agent and a live webpage. A site that implements WebMCP exposes a manifest of callable actions — submitForm(), addToCart(), fetchAccountBalance() — that agents can invoke without scraping HTML or simulating mouse clicks. The agent gets structured data back, not a screenshot.
The practical implications are significant. Right now, browser automation agents (Playwright-based, computer-use models, etc.) are brittle — they break when a site's layout changes. WebMCP would give agents a stable, semantic contract with web applications. Chrome DevTools is also getting an Agent panel in a future update that lets you observe and debug WebMCP calls the same way you'd inspect network requests today.
Whether WebMCP gets traction outside Google's ecosystem is the open question. It needs buy-in from web developers and competing browser vendors. But the proposal is technically thoughtful, and the Hacker News thread is full of developers who've been waiting for exactly this kind of standardization.
What This Means for Your Stack
The shift from assistants to autonomous agents isn't just a product marketing angle — it has real architectural consequences for developers.
- Observability becomes critical. Agents make chains of decisions that are hard to inspect after the fact. Logging, tracing, and replay tooling need to be first-class citizens in any agent-powered system, not an afterthought.
- Auth and permissions get complicated. An agent acting on a user's behalf needs scoped, revocable credentials. OAuth 2.0 works, but the UX of explaining what an agent can and can't do — especially to non-technical users — is still an unsolved problem.
- Cost modeling changes. With subagent orchestration, a single user request can fan out into dozens of model calls. Antigravity 2.0's routing features help, but teams need to instrument agent pipelines with cost tracking from day one.
Google's announcements won't ship to developers overnight — Antigravity 2.0 is in preview, and WebMCP is still a draft spec. But the direction is clear, and teams building developer tools, productivity apps, or anything web-facing should start thinking now about how agents will interact with their surfaces.
The Bottom Line
Google I/O 2026 was the clearest signal yet that the industry has moved past the "AI feature in an existing product" phase and into the "AI as the product" phase. If you're not thinking about how autonomous agents will use your APIs and web surfaces, 2026 is the year to start.