AI DAILY / DEV
←
WEDNESDAY
May 20, 2026
→

    Claude Managed Agents Get Self-Hosted Sandboxes and Private MCP Tunnels
    • Self-hosted sandboxes shipped in public beta — tool execution moves to your infra while Anthropic still runs orchestration, context, and recovery.
    • Launch providers are Cloudflare, Daytona, Modal, and Vercel; each ships a configured guide for the Claude Platform.
    • MCP tunnels in research preview let agents reach private MCP servers over a single outbound connection — no inbound firewall rules, no public endpoints.
    • Lands the day before Anthropic's Code with Claude London Day 1 keynote (May 20).
    tools claude.com

    Google Sunsets Gemini CLI on June 18, Folds It Into Antigravity CLI
    • Gemini CLI and Gemini Code Assist extensions stop serving Google AI Pro, Ultra, and free Code Assist accounts on June 18, 2026.
    • Standard and Enterprise licenseholders keep current access — the cut hits independent devs and hobbyists first.
    • Replacement Antigravity CLI is a ~140 MB Go binary with built-in browser control, sandbox, Git policies, skills runtime, and subagents.
    • Agent Skills, Hooks, Subagents, and Extensions all carry over as Antigravity plugins.
    • HN thread on the front page within hours; commenters bristling at the forced migration and binary size.
    tools developers.googleblog.com

    Gemini 3.5 Flash Lands at 3x the Old Flash Price as Google Aims It at Everything
    • $1.50 / $9 per 1M tokens — 3x Gemini 3 Flash Preview and 6x Gemini 3.1 Flash-Lite.
    • Gets within striking distance of Gemini 3.1 Pro ($2 / $12) and beats it on Terminal-bench 2.1 (76.2%) at 4x the speed.
    • Simon Willison's read: 'Flash' is no longer the cheap tier — Google plans to route everything through 3.5 Flash and price accordingly.
    • Live across the Gemini API, AI Studio, Antigravity, Vertex AI, the Gemini app, and AI Mode in Search; GA on GitHub Copilot the same day.
    models simonwillison.net

    WebMCP Hits Chrome 149 Origin Trial, Letting Sites Expose Tools to Browser Agents
    • Proposed open web standard that turns any page into an MCP server for in-browser AI agents.
    • Two APIs: a declarative one wired to HTML forms, and an imperative one for JS-driven workflows.
    • Replaces brittle DOM scraping with structured, machine-friendly tool calls the page itself defines.
    • Origin trial opens with Chrome 149; available today behind a flag for local development.
    frameworks developer.chrome.com

    Google Punts Gemini 3.5 Pro to June After Audible Groan at the I/O Keynote
    • Pichai on stage: 'I know you can't wait to get your hands on it. Give us until next month.'
    • Google says 3.5 Pro is already running internally; the delay is to RL-train it on production traffic from 3.5 Flash first.
    • Means the day-one I/O frontier story leans entirely on Gemini 3.5 Flash plus Omni for video.
    • Pro is still expected to land behind Anthropic's Mythos and OpenAI's GPT-5.5 on the hardest agentic-coding evals.
    models dnyuz.com

    Karpathy's autoresearch Loop Spreads as Shopify's 53% Speed PR Gets Flagged Overfit
    • Single-file ML experiment harness Karpathy open-sourced in March; ~25.9K stars and +1.7K today on GitHub Trending.
    • Shopify filed a 93-commit PR claiming a 53% speedup on the nanochat training baseline.
    • Reviewers flag the result as overfit to the harness's scalar metric — the PR is still unmerged.
    • Signal: agents are now cheap enough that teams optimize overnight until the eval breaks rather than the model.
    open-source github.com