Google I/O 2026: Gemini 3.5 Flash, Antigravity 2.0, and Managed Agents in the API
Google I/O 2026 marked a deliberate pivot from "AI assistants" to "autonomous agents," and the developer announcements were structured around that transition. The event introduced several new products and APIs that collectively define what Google expects developer workflows to look like in the agentic era.
Gemini 3.5 Flash is the new default model for most developer use cases. Google positioned it as the best-in-class for coding and long-horizon agentic tasks at a cost and latency profile suitable for production applications. On benchmarks it significantly outperforms the previous Gemini 2.5 series while maintaining the sub-second response times that make agentic loops practical. Alongside Flash, Google unveiled Gemini Spark, a beta AI agent built on Gemini's reasoning capabilities that runs persistently in the background, connects to a user's apps (Google Workspace, third-party services), and takes autonomous action without requiring per-task invocation. Spark represents Google's answer to the "always-on agent" pattern that Claude Code and Codex introduced for development workflows, extended to general productivity. A Google AI Ultra subscription tier at $100/month provides priority access to Gemini flagship models and Spark.
Antigravity 2.0 is the platform-level rebranding and expansion of Google's multi-agent development infrastructure. The headline addition is a new CLI that allows developers to define, configure, and deploy specialised subagents, with built-in sandboxing for agent execution and hardened Git policies for agents that modify repositories. Managed Agents in the Gemini API are the most directly developer-relevant feature: they provision isolated remote Linux sandboxes that an agent can use to plan and execute code, handle file I/O, and run terminal commands — removing the burden of building and securing agent execution environments from application developers. The API handles the infrastructure; developers provide the task and tools.
For Android developers, a Migration Agent automates the conversion of React Native or iOS codebases to native Kotlin, using Gemini to understand the source structure and emit idiomatic Android code. The WebMCP proposal is worth watching longer-term: it is Google's bid to standardise how browser-based agents (Chrome extensions, web automation tools, IDE-to-browser workflows) communicate with web applications, analogous to how MCP standardised agent-to-tool communication in desktop environments.
Read more — Google Developers Blog
Read more — Google Blog
GitHub Copilot Removes Gemini Models from Web Interface
GitHub published a changelog update on May 20 announcing that all Gemini models have been removed from Copilot Chat on the web, alongside several GPT-5 variants including GPT-5.2 Codex and GPT-5.4 Nano. The remaining model roster focuses on OpenAI's GPT-5 series and Anthropic's Claude models.
The stated rationale is response consistency: GitHub has observed that the web Copilot experience produces more reliable results with a smaller, curated set of models rather than a broad picker that includes models optimised for different use cases or latency profiles. From a strategic perspective, this is also notable timing — it follows Google I/O by one day and removes Google's models from Microsoft's developer platform just as Google made significant developer announcements.
For developers using Copilot Chat on the web for coding assistance, the practical impact depends on which model they have been using. Teams that prefer a specific Gemini variant will need to evaluate an OpenAI or Claude alternative. The GitHub mobile app and IDE extensions are not affected by this change; the removal is specific to the web interface's model picker.
Read more — GitHub
Links & Sources
You May Also Enjoy
