I/O Day 2: 85+ Sessions, Every Deep-Dive Ranked

This post anchors to Google I/O 2026 Day 2 (May 20, 2026) — the developer-track day that pairs each Day 1 keynote announcement with deep-dive sessions and codelabs, with 85+ on-demand sessions going live tomorrow (May 21). If you only have 90 minutes, this curated matrix tells you exactly which sessions to watch first, in which order, and why each one matters for your production stack.

Yesterday (May 19) Google announced Gemini 3.5 Flash GA, Antigravity 2.0 with a standalone desktop app and CLI, Managed Agents via the Interactions API, Android 17 "Cinnamon Bun," Firebase Genkit 2.0, and the A2A protocol backed by 150+ organisations. Today those launches became hands-on developer sessions. The complete Day 1 announcement index lives in our I/O 2026 complete announcement guide — this post is the Day 2 developer-track companion.

What follows is the only published watch-priority matrix that maps each Day 1 keynote product directly to the Day 2 session that goes deep on its implementation. For the API-level gotcha that will silently degrade production systems upgrading to Gemini 3.5 Flash — the thinking_level default flip — keep reading to section 04.

Google I/O 2026 ran May 19-20 at Shoreline Amphitheatre, Mountain View. Day 1 (May 19) was the Google Keynote at 10:00 AM PT plus the Developer Keynote at 1:30 PM PT. Day 2 (May 20) was the full developer session block — tracked sessions, parallel tracks, codelabs — livestreamed through the day. Starting May 21, the complete 85+-session catalogue becomes on-demand at io.google/2026/sessions and the Google for Developers YouTube channel. No login, no ticket, no cost.

The practical implication: you do not need to have been watching live to access every session. The catalogue is yours to curate at your own pace. Sessions can be saved to a personal schedule in advance on the I/O site. The four confirmed Day 2 morning-block sessions began at 10:00 AM PT on May 20 — Flutter, Web UI, Gemma open models, and Android development tools ran in parallel tracks. "Adaptive Development for the Expanding Android Ecosystem" followed in the next block.

Google's session index at io.google/2026/sessions is alphabetical. This matrix ranks the ten highest-value Day 2 sessions by watch priority, maps each to its Day 1 keynote pairing, and gives you a one-line outcome statement so you can decide in 30 seconds whether it belongs in your 90-minute budget. Session titles below reflect those announced or confirmed in Google's developer keynote round-up; verify exact titles and speaker names against the official io.google/2026/sessions catalogue on May 21.

If 90 minutes is your hard constraint: sessions 1-3 are mandatory for any team running Gemini API or building with Antigravity. Session 4 is mandatory for Android developers targeting AI-aware app capabilities. Sessions 5-6 cover the Firebase and Android Studio ecosystem changes that will affect your toolchain within the quarter. Sessions 7-10 are valuable context but can wait for weekend viewing.

Antigravity 2.0 launched yesterday (May 19) — not an incremental update to Antigravity 1.0's VS Code fork baseline, but a fully separate platform: standalone desktop app, the agy CLI, and now the Antigravity SDK that gives developers programmatic control over the agent harness. Day 2 sessions paired the May 19 announcement with deep implementation guidance. The TechCrunch launch recap confirmed that Antigravity 2.0 and its CLI are now two powerful developer surfaces.

The agy CLI is a Go binary with zero runtime dependencies — installation is a one-line curl or PowerShell script, with an OAuth browser flow on first run and a terminal-URL fallback for CI and headless environments. This matters for teams that want to embed Antigravity in CI pipelines without managing a Node or Python runtime dependency. The SDK, meanwhile, gives developers the same programmatic surface as the desktop app — full agent customization and deployment to their own infrastructure.

The configuration convention changed with 2.0. Where Gemini CLI used GEMINI.md, Antigravity 2.0 uses AGENTS.md for project-level instructions, and the skills directory migrates from .gemini/skills/ to .agents/skills/. Any team that was using the Gemini CLI with a GEMINI.md convention will need to migrate both the filename and the directory before their skills activate in Antigravity 2.0.

For a deep lineage comparison against Antigravity 1.0 — the VS Code fork with Manager View and five parallel agents — see our Antigravity 1.0 on Gemini 3 Pro guide. The 2.0 platform is a different product category, not a feature update. Teams evaluating Antigravity for the first time should start with 2.0 and the official Day 2 sessions — do not anchor on any 1.0 benchmarks or configuration conventions.

Gemini 3.5 Flash GA'd yesterday with a benchmark profile that reportedly matches large flagship models on several dimensions — Terminal-Bench 2.1 at 76.2%, MCP Atlas at 83.6%, CharXiv Reasoning at 84.2%, and output token generation benchmarks suggesting 4× faster throughput than other frontier models. The pricing story is equally strong: $1.50 / $9.00 per million input/output tokens, roughly 25% below Gemini 3.1 Pro and approximately 3.3× cheaper on input than GPT-5.5 or Opus 4.7. Our Gemini 3.5 Flash benchmarks and API guide covers the full numbers. The Day 2 sessions went deeper on the API-level migration patterns — including the change that will silently degrade any upgrade-blind production system.

The thinking_level parameter replaces thinking_budget. The new field is a string enum with four values: minimal, low, medium, and high. The critical change: the default is medium in Gemini 3.5 Flash GA — down fromhigh in Gemini 3 Flash Preview. The ai.google.dev documentation is the primary source for the enum values.

The benchmark picture rewards a closer read. The 76.2% Terminal-Bench 2.1 figure is a Gemini 3.5 Flash benchmark, not the Antigravity 1.0 SWE-bench number — these are distinct benchmarks measuring different things; conflating them is the most common fabrication in I/O coverage. Google has not published dedicated Antigravity 2.0 SWE-bench numbers; only the 2.0 launch announcement exists as the capability claim anchor. The implication for teams evaluating whether to migrate production Gemini API calls: run your own prompts against the actual enum values before committing. The default shift from high to medium can be corrected with a single parameter — but only if you know to set it.

Managed Agents in the Gemini API launched May 19 as the answer to a perennial developer complaint: the scaffolding tax. Setting up a sandboxed Linux environment for an agent to reason, execute code, browse the web, and manage files has historically required significant infrastructure work — Cloud Run configuration, Docker container management, network policy decisions. The Interactions API removes that friction with a single API call that provisions the entire remote execution environment. The official Interactions API documentation and code patterns are now live at ai.google.dev.

The Managed Agents surface is accessible via Google AI Studio and the Gemini Enterprise Agent Platform. Day 2 sessions covered the Interactions API code patterns — how to structure agent tasks, handle tool calls, manage state across multi-step execution, and observe agent behavior through the built-in logging. For teams comparing this to running a self-managed agent on Cloud Run: Managed Agents is the right choice when you want infrastructure abstracted away and are comfortable with the Google-managed execution environment; Cloud Run with your own container is still the right choice for sovereignty-sensitive workloads, custom networking, or fine-grained cost controls.

The lineage here connects to two predecessor surfaces: Jules, Google's async coding agent, explored the managed-execution model for async coding tasks. Managed Agents in the Interactions API generalises that pattern to arbitrary agent workloads via a developer-facing API — a meaningful step up in surface area.

Android 17 "Cinnamon Bun" is targeting a stable release in June 2026. The Android Developers Blog's "17 things for Android developers" post is the primary technical source — it covers everything that mattered in the Android Day 2 track. For AI app developers, two changes are structural: AppFunctions and Jetpack Compose becoming the official UI standard.

AppFunctions is the new Android platform API — with an accompanying Jetpack library — that allows apps to expose their capabilities as MCP servers for AI agents. An app that implements AppFunctions effectively becomes a tool that any AI agent (including Gemini) can discover and invoke via the MCP protocol. Integration with Gemini is currently in private preview; Early Access Program sign-ups were announced at I/O. This is the most consequential long-term platform change in the Android 17 stack — every app that implements AppFunctions becomes part of the agent-accessible tool ecosystem.

Jetpack Compose is now Android's official UI standard. Views are in maintenance mode after a five-year migration period — future Android guidance and libraries will be Compose-first by default. Jetpack Glance for Widgets ships in Android 17 with RemoteCompose, unifying widget rendering across mobile, Wear OS, and automotive. The Compose consolidation means any team still writing new Views code is accumulating technical debt.

Performance engineers got Lock-free MessageQueue and refined garbage collection. Media teams get CameraXViewfinder Composable and the Media3 AI Effects library (Magic Eraser, Studio Sound). Teams targeting codec-specific encoding have CodecDB for chipset-specific decisions. The Migration Assistant in Android Studio can now port iOS, React Native, or web framework apps to native Kotlin with Jetpack Compose — "in hours instead of weeks," according to the Android Developers Blog.

Gemini Nano 4 preview is live for on-device tasks — data extraction, summarization, and structured output. Firebase AI Logic now supports Gemini Nano 4 hybrid inference on iOS and Gemma 4 on Android. The on-device agent stack adds the Agent Development Kit (ADK) and new AG-UI / A2UI communication protocols. For teams building AI-aware Android apps, this is the session to watch before any architectural decisions around on-device vs cloud inference.

Firebase Genkit 2.0 is now generally available. The Firebase blog's I/O 2026 announcement post covers the full changelog. The headline additions: native MCP server integration, streaming support, improved observability via Cloud Trace integration, local development with the Firebase emulator, and multi-model routing. Genkit 2.0 is Google's TypeScript and JavaScript AI application framework — if you are building AI features in a Firebase-hosted app, Genkit 2.0 is the Day 2 session most likely to affect your next sprint.

Firebase AI Logic (previously Firebase ML) received several AI Studio-adjacent updates: grounding with Google Maps to reduce hallucinations in location-aware applications; "template-only" and "authentication-mode" security policies; hybrid inference for iOS with Gemini Nano 4; Gemma 4 support for Android on-device inference; and local web inference in Chrome graduating from origin trial to general availability.

Google AI Studio added native Kotlin support for building Android apps directly from a prompt, Google Workspace integrations, one-click deployment to Cloud Run with Firebase service support, and seamless project export to Antigravity for local development. The Firebase → AI Studio → Antigravity pipeline is now a single continuous developer flow — the session pairing for this section covers the end-to-end implementation. Our AI transformation practice can help teams design the right Firebase + Gemini API architecture for their workload.

Three open standards tracks ran through Day 2 that individually look incremental but together represent a deliberate architectural play by Google around agent interoperability.

WebMCP is a proposed open standard announced at the I/O Developer Keynote that enables browser-based AI agents to execute complex tasks with greater speed, reliability, and precision. It pairs with Chrome DevTools for Agents, which automates quality audits and real-world experience emulation. A spec URL was not published alongside the announcement — verify the specification reference against the official Developer Keynote round-up on May 21 when the on-demand content is live.

Chrome built-in AI APIs expanded at I/O 2026. The stable tier (live since Chrome 138 from I/O 2025) includes the Translator, Summarizer, Language Detector, and Prompt APIs for extensions. The 2026 additions include a Writer API and Rewriter API in origin trial, a multimodal Prompt API in Early Preview, and a Proofreader API also in Early Preview. Local web inference in Chrome graduated to general availability via Firebase AI Logic. The HTML-in-Canvas API is in origin trial, enabling searchable, accessible 3D experiences.

A2A (Agent-to-Agent) protocol is the strategically significant headline. MCP handles agent-to-tool connections; A2A handles agent-to-agent task delegation — a different problem at a higher abstraction level. The technical stack: HTTP + Server-Sent Events + JSON-RPC 2.0 transport with Agent Cards for capability advertisement and discovery. The 150+ organisation coalition backing A2A — Microsoft, AWS, Salesforce, SAP, LangGraph, CrewAI, AutoGen, Semantic Kernel, and well over a hundred more — means this protocol has enough ecosystem momentum to matter for any team building multi-agent systems in 2026.

The A2A + MCP combination is what makes AppFunctions architecturally interesting: an Android app becomes an MCP tool server via AppFunctions, discoverable and invokable by any A2A-capable agent in the ecosystem. The Day 2 sessions on WebMCP and A2A covered the protocol details — both specs are young enough that implementation guidance will evolve quickly over the summer.

The Day 1 keynote launched products; Day 2 sessions shipped implementation depth. This crosswalk maps each major May 19 announcement to the Day 2 session that goes deep on its API, SDK, or migration pattern. Session titles flagged with [verify] should be confirmed against io.google/2026/sessions on May 21 — the official session index was not fully machine-readable at research time.

The complete Day 1 announcement index — all 100+ things Google announced at I/O 2026, from AI Mode's 1B monthly active users to the Gemini Spark / Omni / Universal Cart launches — lives in our I/O 2026 complete announcement guide. Use that post as the Day 1 reference; use this post as the Day 2 session shortcut. The two posts are designed to work together.

From an analytical standpoint, the Day 1 → Day 2 mapping reveals Google's strategic priority ordering more clearly than the keynote itself. The sessions with the most depth — Antigravity SDK, Interactions API, Android AppFunctions — are precisely the surfaces where Google is betting on developer adoption as the growth lever for the Gemini ecosystem. The A2A protocol's 150+ org coalition is the forward-looking bet: if agent-to-agent task delegation becomes a standard protocol the way HTTP became the standard for web requests, Google's positioning as the protocol proposer is a strategic moat that compounds over years. Developers who understand the A2A stack today will be architecting multi-agent systems on a standard that is already broadly endorsed.