Key Signals

• Google sunsets Gemini CLI and transitions all free/individual users to Antigravity CLI by June 18, 2026. Announced at I/O 2026, this unifies Google's terminal AI tools into a single agent-first platform built in Go with asynchronous multi-agent orchestration, shared architecture with the Antigravity 2.0 desktop app, and faster execution. Enterprise customers on paid Gemini Code Assist licenses retain existing Gemini CLI access unchanged. This is the most significant consolidation in the AI coding CLI space this year, eliminating a standalone tool with 100K+ GitHub stars in favor of a unified platform strategy. [1][2]

• Gemini 3.5 Flash launches as a Flash-tier model that rivals frontier competitors at 4x their speed and less than half their price. Scoring 76.2% on TerminalBench 2.1 (vs. 3.1 Pro's 70.3%), it outperforms Google's own Pro model and competes with GPT-5.5 and Opus 4.7 in tool-use benchmarks while outputting ~280 tokens/second. Google co-developed it with Antigravity for native agentic environments, and it is now available in GitHub Copilot with a 14X premium request multiplier. The forthcoming 3.5 Pro will serve as the orchestrator model while Flash handles sub-agent brute-force tool use. [3][4][5]

• Anthropic ships MCP tunnels and self-hosted sandboxes for Claude Managed Agents. MCP tunnels let agents connect to private MCP servers inside enterprise networks via a lightweight outbound-only gateway — no inbound firewall rules needed. Self-hosted sandboxes move tool execution to customer infrastructure through providers like Cloudflare, Daytona, Modal, and Vercel while Anthropic retains agent-loop orchestration. This directly addresses the enterprise bottleneck where security reviews block production agent deployments. [6][7]

• Copilot CLI v1.0.49 delivers a major feature release including /rubber-duck for independent agent critique, persistent /memory controls, and experimental MCP server discovery. The release adds Alpine Linux support, /chronicle search across all session content, plugin bulk updates, and properly fires hooks for sub-agent tool calls. It also introduces a "None" reasoning effort option and COPILOT_PLUGIN_DIR_ONLY for deterministic plugin sets — reflecting growing demand for reproducible agentic workflows. [8]

• Claude Code v2.1.144 makes background sessions first-class with /resume support, elapsed duration tracking, and dozens of stability fixes. The release addresses a critical startup hang (up to 75s) when api.anthropic.com is unreachable, fixes MCP servers with paginated tools/list silently dropping tools, and improves SDK/headless MCP startup by up to 2 seconds. Per-session model switching via /model and the rename from "extra usage" to "usage credits" signal continued maturation of Claude Code's multi-session workflow. [9]

• Google opens Android Studio to GPT, Claude, and local Gemma 4 models — embracing model plurality for AI-assisted mobile development. Combined with the Android CLI reaching 1.0 stable, and the proposal of WebMCP as an open standard for browser-based AI agents, Google is positioning its developer tools as model-agnostic infrastructure rather than Gemini-exclusive walled gardens. Chrome DevTools for agents also reached 1.0, giving 20+ coding agents direct access to console logs, network traffic, and accessibility trees. [10][11]

AI Coding News

• Anthropic debuts MCP tunnels and self-hosted sandboxes at its Code with Claude conference in London, separating agent orchestration from execution for enterprise deployments. Self-hosted sandboxes in public beta allow tool execution on customer-controlled infrastructure via Cloudflare, Daytona, Modal, and Vercel. MCP tunnels in research preview deploy a lightweight gateway making a single outbound encrypted connection, managed from Claude Console workspace settings. Clay uses it for its Sculptor engineering agent, DoorDash is building internal productivity agents on Modal, and Rogo runs financial AI workloads on Vercel. Neither feature requires changes to existing Claude Managed Agents integrations. [6][7]

• Google proposes WebMCP as an open web standard and ships Chrome DevTools for agents 1.0 to make the web agent-ready. WebMCP allows developers to expose JavaScript functions and HTML forms to browser-based AI agents, enabling them to interact with sites programmatically rather than via screenshots or DOM traversal. An origin trial begins in Chrome 149 with Booking.com, Expedia, Instacart, Shopify, and Redfin showing early interest. Chrome DevTools for agents provides direct access to console logs, network traffic, and accessibility trees via a built-in MCP server or CLI. Modern Web Guidance launches in early preview as a skill set for coding agents, installable via npx modern-web-guidance install, teaching agents about Baseline browser support targets with fallback strategies. [11]

• Google's I/O 2026 strategic bet frames Gemini 3.5 Flash as the engine for autonomous multi-hour coding agents, not conversational chatbots. DeepMind CTO Koray Kavukcuoglu confirmed an optimized Flash variant runs 12x faster than frontier models at equivalent quality. At the conference, engineers demonstrated agents spawning sub-agents to build a full operating system inside Antigravity. Google's product lead Tulsee Doshi outlined the intended architecture: 3.5 Pro as orchestrator/planner, Flash as parallel sub-agent workers handling brute-force tool use. The model can run autonomously for multiple hours, pausing only at decision points requiring human judgment. [3]

• Google AI Studio now lets anyone build native Kotlin Android apps from prompts with a built-in emulator, device testing via ADB, and direct Play Console upload. This extends vibe coding from web-only workflows to full mobile development without installing Android Studio. Apps use Jetpack Compose and can access device sensors. Google plans Firebase integrations for databases and authentication, plus direct sharing from AI Studio. A Google Play developer account ($25 one-time fee) enables uploading to internal testing tracks. [16]

Feature Update

• GitHub Copilot renames "Implement suggestion" to "Fix with Copilot" and adds batch processing for code review feedback. The new dialog lets users choose between applying changes directly to the current PR or opening a new PR, select the model for implementation, and add optional instructions. The "Fix batch with Copilot" button in the PR Overview comment enables batching multiple review comments before handing them to Copilot cloud agent, eliminating the need to handle each suggestion individually. [12]

• Gemini 3.5 Flash is generally available in GitHub Copilot across VS Code 1.115+, Visual Studio 17.14.22+/18.1.0+, JetBrains, Xcode, and Eclipse. The model launches with a tentative 14X premium request multiplier and is available to Copilot Pro, Pro+, Business, and Enterprise users. Copilot Business and Enterprise administrators must explicitly enable the Gemini 3.5 Flash policy in settings. Google positions it as delivering near-Pro coding quality at Flash-tier speed with high cache efficiency for iterative agentic workflows. [5]

• Copilot CLI v1.0.49 ships a feature-dense release with 12 new capabilities and 17 bug fixes. Highlights include /rubber-duck for independent critique of the agent's work, /chronicle search across all session content, /memory on|off|show for persistent memory control, /session id to display and copy session IDs, Alpine Linux support, and experimental /mcp search to discover and install MCP servers from registry. Hooks now fire correctly for sub-agent tool calls, and PromptFrame UI renders inside tmux with ghostty, WezTerm, or kitty terminals. [8]

• Copilot CLI v1.0.51-1 ships as a pre-release on May 19, preceding v1.0.51-2 which adds /memory show documentation links and fixes session naming for usage-based billing users. This pre-release is part of the rapid iteration cadence between the v1.0.49 stable release and the next stable version. [13]

• Claude Code v2.1.144 adds /resume support for background sessions and per-session model switching. Background sessions started via claude --bg or agent view now appear alongside interactive ones. The /model command changes models for the current session only, with d in the picker setting defaults for new sessions. Critical fixes include startup no longer hanging 75s when api.anthropic.com is unreachable (15s timeout), MCP servers with paginated tools/list no longer silently dropping tools, and macOS background sessions no longer crashing under Full Disk Access-protected folders. SDK/headless MCP startup is up to 2s faster by overlapping pre-wait with initialization. [9]

• Cursor launches a Jira integration enabling cloud agent task execution directly from work items. Users can assign Jira tickets to Cursor or mention @Cursor in comments to start a cloud agent that uses the ticket title, description, comments, and repository settings as context. The agent can fix bugs, add features, update tests, or investigate issues, posting completion updates and PR links back to Jira. Requires Cursor admin access and Jira Commercial Cloud with Rovo enabled. [14]

• Kiro ships session stability improvements with stop control and mobile layout fixes. Users can now stop the agent mid-task, see progress while workspaces spin up, and benefit from a polished layout on smaller screens. The release resolves several session stability issues in the web interface. [15]

• Google officially transitions Gemini CLI to Antigravity CLI, available immediately, with Gemini CLI sunsetting June 18 for free users. Built in Go for faster execution, Antigravity CLI supports asynchronous multi-agent workflows and shares the same agent harness as Antigravity 2.0 desktop. Core features preserved include Agent Skills, Hooks, Subagents, and Extensions. Gemini Code Assist for GitHub will also stop accepting new installations on June 18. Enterprise customers on paid licenses are unaffected and retain Gemini CLI access with latest models. [1]

• Google I/O 2026 Developer keynote introduces Antigravity 2.0, Antigravity SDK, Android CLI 1.0, and Managed Agents API. Antigravity 2.0 is a standalone desktop app with cross-platform terminal sandboxing, credential masking, and hardened Git policies. The Antigravity SDK provides programmatic control over the agent harness for custom deployment. Android CLI 1.0 enables any agent to use Android Studio capabilities. Managed Agents in the Gemini API delivers a fully provisioned agent with remote sandbox via single API call. A Migration agent can convert React Native, iOS, or web apps to native Kotlin in hours instead of weeks. [2]

• Gemini 3.5 Flash scores 76.2% on TerminalBench 2.1 and 83.6% on MCP Atlas, outperforming 3.1 Pro and competing with GPT-5.5 and Opus 4.7 at ~280 tokens/second. Available via Gemini API in Google AI Studio, Android Studio, Vertex AI, Gemini Enterprise, and Antigravity. Artificial Analysis positions it just behind frontier models in quality but at dramatically faster throughput and less than half the price. A 3.5 Pro model is expected next month. [4]

• Android Studio now supports GPT from OpenAI, Claude from Anthropic, and local Gemma 4 alongside Gemini models. Developers can download Gemma 4 directly from the latest canary build without an external server. The Android CLI 1.0 introduces android studio commands enabling agents to perform semantic symbol resolution, analyze warnings, and render Jetpack Compose previews. Agent Skills ship built-in for Android development and Firebase. On Android Bench, GPT 5.5 currently leads with ~75% test case resolution, followed by GPT 5.4 and Gemini 3.1 Pro Preview. [10]