Full sweep this week. I considered being brief, but the news didn't cooperate.
Claude Code now offers routines in research preview — saved configurations (a prompt, a repo, and your connectors) that run on Anthropic's cloud infrastructure on a schedule, from an API call, or in response to GitHub events like new PRs or pushes. Your laptop can be closed. Use cases range from nightly backlog grooming and bespoke code review on every PR to deploy verification and cross-repo library ports. This is Claude Code's clearest step yet from developer tool to autonomous infrastructure. Claude Code Docs
Claude Code v2.1.92–v2.1.101 added four features worth your attention: Ultraplan, which starts a cloud-based planning session from the terminal that you can annotate in the browser and execute remotely; the Monitor tool, which spawns background watchers that stream events into your conversation — tail logs, babysit CI, auto-fix crashes; /autofix-pr for fixing PRs from the CLI; and /team-onboarding for generating ramp-up guides from your actual usage. The /loop command now self-paces instead of requiring a fixed interval. Claude Code continues its quiet migration from interactive prompt to ambient infrastructure. Claude Code What's New
Anthropic's new advisor tool lets developers pair a fast executor model (Sonnet or Haiku) with Opus as a strategic advisor that steps in at complex decision points. Haiku with an Opus advisor more than doubled its BrowseComp score while costing 85% less per task than running Opus alone. If you've been choosing between cost and capability on agentic workloads, the advisor pattern makes that tradeoff significantly less painful. Anthropic Release Notes
Anthropic's latest model, Claude Mythos Preview, autonomously discovered thousands of zero-day vulnerabilities across every major OS and browser — including a 27-year-old OpenBSD kernel bug. Rather than release it publicly, Anthropic restricted access to 12 security partners (AWS, Apple, Google, Microsoft, CrowdStrike, and others) under Project Glasswing, committing $100M in compute credits for defensive security work. It scores 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro. A voluntary capability restriction of this magnitude is new territory. TechCrunch The Verge
OpenAI's Chief Revenue Officer circulated an internal memo stating the Microsoft relationship has "limited our ability to meet enterprises where they are — for many that's Bedrock." Enterprise demand through AWS has been described as "staggering" since Amazon's $50B investment. Microsoft says the partnership remains "strong and central." PYMNTS
Nineteen Markdown workflows encoding a six-phase engineering lifecycle (Define, Plan, Build, Verify, Review, Ship) with verification gates that prevent AI agents from skipping steps like testing. Works with Claude Code, Cursor, Gemini CLI, Windsurf, and Copilot. Already at 14.7K GitHub stars. GitHub
Google added persistent project hubs to Gemini — organise conversations, files, and custom instructions in one workspace, with automatic sync to NotebookLM. Rolling out to Ultra, Pro, and Plus on web. The Gemini-NotebookLM workflow integration is becoming harder to dismiss. TechRepublic
Tariffs on imported semiconductor manufacturing equipment from ASML, Applied Materials, Lam Research, and KLA went live April 9. TSMC announced 5–10% price hikes on sub-5nm wafers globally, with US-fab chips expected to carry an additional 10–30% premium. If you're budgeting for GPU-heavy infrastructure, the cost floor just moved up. Chosun Ilbo
Anthropic's fully hosted orchestration layer for running Claude as an autonomous agent — sandboxed execution, state management, credential handling, multi-agent coordination — is now in public beta at $0.08/session-hour plus tokens. Early adopters include Notion, Asana, and Sentry. Anthropic Release Notes
Bugbot extracts patterns from PR feedback — reactions, replies, human reviewer comments — and generates rules that apply to future reviews. Over 110K repos have enabled learning, resolution rate climbed to 78%, and MCP server support shipped alongside a batch "Fix All" action. Cursor Blog
Meta's newly formed Superintelligence Labs shipped its first model — Muse Spark — through meta.ai and the Meta AI app, with WhatsApp, Instagram, and Facebook rollout coming. The model combines text and image generation with tool use and a multi-agent "Contemplating Mode." It's the first product-facing output from the lab Meta created after the $14.3B Scale AI deal. CNBC
HBM4 memory from SK Hynix and Micron failed qualification, pushing Rubin mass production from June to September 2026. Production target dropped from 2M to 1.5M units, and Rubin Ultra was scaled back from four-die to dual-die due to TSMC packaging yields. Blackwell production is increasing to fill the gap. The Register
Connect Azure OpenAI, Anthropic, Ollama, vLLM, or any OpenAI-compatible endpoint. Set COPILOT_OFFLINE=true for air-gapped development with local models. GitHub auth optional when using your own provider. GitHub Changelog
Anthropic secured 3.5 gigawatts of next-generation Google TPU capacity through Broadcom starting 2027, on top of 1 GW already being delivered. Revenue run rate surpassed $30B — up from $9B at end of 2025. Enterprise customers spending over $1M annually doubled to 1,000+ in under two months. Anthropic
GitHub Copilot CLI now includes Rubber Duck, an experimental feature that uses a model from a different AI family to review the primary agent's plans and work. Claude Sonnet paired with GPT-5.4 as Rubber Duck closes 74.7% of the Sonnet-to-Opus performance gap on SWE-Bench Pro, with the biggest gains on complex multi-file tasks. GitHub Blog
After Trump announced the rescue of a downed F-15 crew, AI-generated images went viral — shared by Texas Gov. Greg Abbott, AG Ken Paxton, and NY Rep. Mike Lawler before community notes intervened. One post reached 791K views. The "politician sharing deepfakes during a live military operation" milestone has been reached. The Guardian Military.com
That's the lot. If anything on this list surprised you, you're not reading these often enough.