Good morning. It’s Wednesday, April 14th.

You read. We listen. Let us know what you think by replying to this email.

^{In partnership with TinyFish}

AI agents like OpenClaw and Claude Code are powerful. But they can't use the web besides basic search and fetch.

For anything beyond that, they'll spin up Playwright scripts, dump entire raw HTML pages into their context, or hit anti-bot walls and silently fail. You end up maintaining pipelines instead of shipping.

TinyFish just launched the full-stack web infra for AI agents with Search, Fetch, Browser, and Agent primitives - all built in-house. All under one API key.

And the fastest way in is their CLI + Skill: install the CLI (npm install -g @tiny-fish/cli), drop a single markdown Skill file into your agent's directory, that’s it!

• 87% fewer tokens than MCP per operation
• #1 web agent on Mind2Web (89.9% accuracy)
• Run at enterprise scale with parallel browser sessions
• Works with OpenClaw, Claude Code, Cursor, OpenCode, ad other agents

^{Thank you for supporting our sponsors!}

Anthropic just overhauled its Claude Code desktop app into a mission control for parallel agents, allowing developers to orchestrate multiple refactors and bug fixes across repos simultaneously.

To solidify this as a persistent automation layer, Anthropic also launched ‘routines,’ a cloud-based, event-driven workflows that handle tasks like PR triage and automated reviews via GitHub hooks without needing a local machine. However, some users are flagging degraded reasoning depth, a likely symptom of Anthropic throttling effort settings to manage GPU shortages.

Co-founder Jack Clark confirmed the company recently briefed the Trump administration on Mythos, a high-risk model tested by the AI Security Institute. In controlled AISI evaluations, Mythos completed a 32-step enterprise cyberattack simulation in 3/10 attempts, averaging 22/32 steps, highlighting dual-use cyber capability and the urgency of stronger defensive security practices.

Anthropic is also eyeing the creative market. The upcoming Claude Opus 4.7 web design tool is already rattling SaaS stocks, while Opus 4.6 agents recently achieved a 0.97 performance gap recovery in automated alignment research. At the same time, leaked prototypes point to a Claude in-chat app builder that generates full applications from prompts, putting it in the same space as Europe’s Lovable and the wider no-code ecosystem. Read more.

OpenAI is doubling down on a platform-first strategy to box out Anthropic, according to a leaked Denise Dresser memo. The centerpieces are ‘Spud,’ a new model tuned for reasoning and reliability, and ‘Frontier,’ a platform designed to let autonomous agents loose on enterprise workflows. By plugging GPT-5.4 and Codex into Cloudflare’s edge, OpenAI is moving agents out of the chat box and into real-world system automation.

To lock in defenders, the company launched GPT-5.4-Cyber under an expanded "Trusted Access" program, giving vetted pros less restrictive tools for binary analysis. It’s a direct response to Anthropic’s Mythos. Greg Brockman is pitching this as the dawn of a compute-powered economy where human intent is the only bottleneck, and small teams use massive compute to act like giants.

OpenAI is expanding its London office to 500+ staff despite delays to its UK infrastructure plans. It is also backing an Illinois bill that limits liability for large-scale AI harms under defined conditions. It also acquired personal finance startup Hiro. Read more.

The new Google app for Windows brings to the desktop an AI Mode that uses an Alt + Space shortcut to query local files, Drive content, and on-screen data. In the browser, Chrome’s new Skills feature allows users to save prompts as reusable, one-click workflows triggered by slash commands, while Gemini with Personal Intelligence now pulls context from Gmail, Photos, and YouTube to ground responses in user history.

For power users, Google is introducing specialized environments like the Agent tab in Gemini Enterprise, which mimics Claude-style orchestration with task inboxes and human-in-the-loop toggles. AI Studio now features Tab Tab Tab, an autocomplete system that expands fuzzy intent into structured prompts for faster vibe coding.
Research workflows are also getting an upgrade as Google tests Canvas and Connectors in NotebookLM, enabling users to transform sourced data into interactive timelines or apps.

Google DeepMind also released Gemini Robotics-ER 1.6, an embodied AI model that improves spatial reasoning, multi-camera understanding, and task planning using agentic vision with tools like code execution and Search. It also boosts pointing accuracy, task completion detection, and industrial instrument reading, and is available via the Gemini API and AI Studio. Read more.

Watch: Google and Boston Dynamics give robots "agentic vision" to read industrial gauges

Figma for Agents lets you design with AI using your system, so everything stays on-brand while connecting design and code.

Luma Agents plan, create, and refine video, image, and audio together with shared creative context.

Ovren handles your backlog with AI engineers, delivering real code updates you review and approve.

Skills Janitor shows which Claude Code skills you actually use, fixing overlaps, issues, and unused clutter.

Recall 2.0 builds an AI from your knowledge, letting you search, compare, and act across everything you’ve saved.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.