Plenty happened while you were busy. Appropriate to the day, I do find this rather taxing.
Anthropic rebuilt the Claude Code desktop app around running multiple agent sessions at once — a sidebar manages active sessions across repos, with drag-and-drop panes for terminal, diff viewer, file editor, and chat. Side chat (Cmd+;) branches off questions without polluting the main thread, and three verbosity modes let you dial from full tool-call transparency to just results. The design philosophy has shifted from "one prompt, one task" to "orchestrator seat with several agents in flight" — which is how most people were already using it, just less comfortably. Claude Blog
User frustration with perceived Claude performance degradation crossed from GitHub issues into Fortune and The Register. Developers report weaker sustained reasoning, mid-task abandonment, and token waste — some characterising it as "AI shrinkflation." Anthropic denies intentional degradation but acknowledged changes to usage limits and reasoning defaults. SWE-Bench Pro scores show no measurable decline, which hasn't settled the argument. Fortune The Register
Save a Gemini prompt as a reusable "Skill" — one click runs it across whatever page or tabs you're viewing. Google shipped 50+ pre-built Skills for research, shopping, and productivity, and you can create custom ones from chat history. Access via / or the + button in the Gemini sidebar. Rolling out to Chrome desktop (English US) first. A lightweight automation layer for the browser, if you're already in the Google ecosystem. TechCrunch
Week 15 brought four features worth your attention: Ultraplan offloads planning to a cloud session you can annotate in the browser while keeping your terminal free. The Monitor tool spawns background watchers that stream CI logs, deployment events, and crash reports into your conversation. /autofix-pr enables PR auto-fix from the CLI, and /team-onboarding generates ramp-up guides from your actual usage patterns. Claude Code's quiet migration from interactive prompt to ambient infrastructure continues at pace. Claude Code What's New
In V8.0, generating native 2K images cost 4x the GPU time of a standard image, which burned through subscription hours fast. V8.1 makes HD 3x cheaper and 3x faster, enough that it's now the default — every generation is 2K unless you opt out. Standard resolution also got 25% cheaper and 50% faster, matching V7 draft speed. Image prompts and image weights return after being absent in V8.0. Midjourney Updates
Notion ships major releases in numbered instalments called "parts" — and yes, they're really calling it that. Custom Agents are now 35–50% cheaper to run, with new budget models (GPT-5.4 Mini & Nano, Haiku 4.5) using up to 10x fewer credits. AI Autofill brings agent-powered enrichment directly into databases — continuously categorise, extract, and update rows without manual review. The Notion Agent gains saveable Skills for repeatable workflows, plus integrations with Calendar, Mail, and Slack. New connectors for Salesforce and Box round it out. Custom Agents remain free until May 3. Notion Releases
Z.AI released GLM-5.1, a 754-billion parameter mixture-of-experts model under MIT license. It scores 58.4 on SWE-Bench Pro — beating GPT-5.4 (57.7), Claude Opus 4.6 (57.3), and Gemini 3.1 Pro (54.2). Built for sustained agentic execution up to 8 hours, it also leads on CyberGym (68.7) and Terminal-Bench 2.0 (63.5). The open-weight self-hosted option just became significantly more competitive with the frontier. MarkTechPost Hugging Face
The BBC's Top Comment podcast identified "Mr Explosive" of Explosive Media, who confirmed the Iranian government is a customer behind the LEGO-style AI propaganda videos about the US-Iran conflict. The videos have been viewed hundreds of millions of times. The LEGO aesthetic deliberately evades content moderation filters designed for realistic military imagery. The term "slopaganda" apparently undersells it. BBC
The expanded agreement runs through December 2032, bringing the total commitment to roughly $35 billion with initial deployments including the NVIDIA Vera Rubin platform. Meta is doubling down on external compute even as it plans $115–135 billion in internal capex for 2026. CNBC
A new API pattern that lets a fast executor model (Haiku 4.5 or Sonnet 4.6) consult Opus 4.6 mid-generation for strategic guidance on complex agentic tasks. The advisor produces a plan, the executor continues — delivering near-advisor quality at executor-model rates. A useful architectural pattern for anyone running long-horizon agents on a budget. Anthropic API Release Notes
Anthropic released a fully managed agent runtime — send an API call, Claude runs autonomously in a sandboxed environment with bash, file operations, web search, and MCP tools, streaming progress back via server-sent events. The new ant CLI wraps the same service for terminal use with YAML-based resource versioning. If you've been stitching together agent harnesses from open-source parts and hope, this is Anthropic's pitch to handle the infrastructure. Whether that trade-off appeals depends on how much you enjoy maintaining your own sandbox. Anthropic API Release Notes
Bugbot turns reviewer reactions, replies, and human comments into learned rules that shape future reviews. Resolution rate has climbed to 78%, and a new "Fix All" action applies multiple suggestions in one operation. MCP server access adds additional review context on Teams and Enterprise plans. An AI code reviewer that improves from being told it's wrong — which is more than can be said for most human reviewers. Cursor Blog Cursor Changelog
Meta Superintelligence Labs' first model is natively multimodal with tool use, visual chain of thought, and multi-agent orchestration. A "Contemplating mode" orchestrates parallel agents to achieve 58% on Humanity's Last Exam. Now powers Meta AI across the Meta AI app, WhatsApp, Instagram, Facebook, Messenger, and AI glasses, with a private API preview for select partners. Meta AI Blog
Persistent project spaces in Gemini that sync bidirectionally with NotebookLM — sources added in one app appear in the other, and you can access NotebookLM features like Video Overviews from within Gemini. Bridges the gap between Gemini's conversational mode and NotebookLM's document analysis. Available to AI Ultra, Pro, and Plus subscribers on web. Google Blog
HBM4 memory qualification issues at SK Hynix and Micron are delaying Vera Rubin GPUs, reducing their share of 2026 high-end shipments from 29% to 22%. Blackwell fills the gap at 71% of shipments. If you were planning around Rubin availability, recalibrate. The Register
Anthropic expanded its TPU partnership with Google and Broadcom to approximately 3.5 gigawatts of capacity arriving in 2027, up from 1 GW last October. The company disclosed a $30 billion annual run rate — up from $9 billion at end of 2025 — with over 1,000 enterprise customers each spending $1M+ annually. Most infrastructure will be U.S.-based as part of a $50 billion domestic commitment. The numbers are starting to sound like defence budgets. TechCrunch Anthropic
Copilot CLI now supports bring-your-own-key providers (Azure OpenAI, Anthropic, any OpenAI-compatible endpoint) and local models via Ollama, vLLM, or Foundry Local. Setting COPILOT_OFFLINE=true enables fully air-gapped operation with all telemetry disabled. GitHub auth becomes optional. Enterprise teams with compliance constraints just gained a significant new option. GitHub Changelog
Nineteen Markdown workflows that encode senior engineering discipline into AI coding agents, covering six phases from Define to Ship. Each skill includes verification gates and anti-rationalization tables that prevent agents from skipping steps like testing. Works with Claude Code, Cursor, Gemini CLI, Windsurf, and Copilot — vendor-agnostic by design. Hit 4,200+ stars in days. GitHub
That's your week. The tools won't stop improving while you read about them, so I'd suggest picking one and getting on with it.