Claude Cowork Breaks Free of the Desktop

🧠 LAUNCH

Claude Cowork Breaks Free of the Desktop.

Claude Cowork is now available on web and mobile, meaning your persistent agentic sessions follow you from laptop to phone to tablet without losing state. Sessions run remotely and persist across devices — no more "I'll pick this up when I'm back at my desk." The Max plan gets it first, but this is the move that turns Cowork from a power-user curiosity into a daily-driver tool. If you've been waiting to try agentic collaboration, the friction just dropped to zero. Read more →

Google Fires Back With Background Agents and Remote MCP in Gemini API.

Google just added background task execution and remote MCP support to its managed agents in the Gemini API — a direct answer to Anthropic's agentic infrastructure push. The significance: production-ready agent hosting is now a multi-provider race, and Google is betting that managed infrastructure (not just model quality) is where the real lock-in happens. If you're building agents, you now have two serious options for hosted runtimes. Read more →

Anthropic Takes Claude Code and Cowork Into Government. Claude Code and Cowork are expanding into the public sector, signaling Anthropic's push beyond commercial enterprise into government workflows. The move follows last week's Alberta cybersecurity case and suggests a systematic FedRAMP-style rollout strategy. Read more →

sqlite-utils 4.0 Ships With Full Schema Migrations. The AI-written library that was RC last week is now production — sqlite-utils 4.0 lands with proper database schema migration support. Simon Willison's $149 experiment has become a real tool with a real release. If you're managing SQLite schemas programmatically, upgrade. Read more →

🔧 TOOL

Claude Code v2.1.203 Polishes the Production Edges: Login expiry warnings, a manual-mode badge, and MCP roots/list improvements — incremental but the kind of polish that matters when you're running Claude Code across a team. The login expiry warning alone will save you from mid-session authentication drops. Read more →

HuggingFace to SageMaker in One Click: HuggingFace now offers one-click deployment from the Hub directly into Amazon SageMaker Studio, eliminating the manual export-configure-deploy dance that turns a 5-minute model evaluation into a 2-hour DevOps exercise. If you're already on AWS, this cuts your model-to-endpoint time dramatically. Read more →

📝 TECHNIQUE

The Official Guide to Picking Your Claude Code Model and Effort Level: Anthropic published the definitive guide on when to use which model and effort setting in Claude Code — Opus for complex multi-file refactors, Sonnet for quick edits, and how effort levels trade latency for accuracy. If you've been running everything on max effort out of habit, you're probably burning 3x the tokens you need. Read more →

🔬 RESEARCH

Lilian Weng Maps the Full Landscape of Self-Improvement Harnesses.

Lilian Weng just published a comprehensive taxonomy of recursive self-improvement — tracing the line from I.J. Good's ultraintelligent machines to modern AI systems that improve their own training loops. The post covers reward hacking, self-play curricula, and the engineering patterns that make self-improvement loops actually converge. This isn't speculation — it's a practical map of techniques that teams are already shipping. Essential reading from one of the field's best technical writers. Read more →

AI Finds Real Bugs in Cloudflare's Cryptography Library. An AI-assisted security audit uncovered genuine bugs in Cloudflare's CIRCL cryptography library — subtle implementation flaws in constant-time operations that human reviewers missed. (70 likes | 9 RTs) This isn't "AI found a typo in a README." These are the kinds of crypto bugs that lead to side-channel attacks, found by an approach that scales to every C library with a fuzzing harness. Read more →

💡 INSIGHT

Early Cowork Adopters Reveal Which Workflows Actually Stick. Anthropic published real usage patterns from early Claude Cowork adopters, and the results are telling — research synthesis and document drafting dominate, while pure coding tasks are less sticky. The pattern: Cowork shines when the task needs persistent context across multiple turns, not when you need a quick function written. Read more →

Latent Space Publishes the Definitive Fable Field Guide. Latent Space consolidates the scattered reactions to Fable 5 into one coherent analysis — what the benchmarks actually mean, where the model excels vs. hype, and how it reshapes the competitive landscape. If you've been piecing together Fable takes from Twitter threads, this saves you three hours. Read more →

The Paradox of Automating AI Away. A sharp essay argues that when your automation layer itself becomes the bottleneck, adding more agents makes things worse, not better. (92 likes | 48 RTs) The core insight: the orchestration overhead of managing AI workflows can exceed the cost of the work being automated. A necessary counterpoint to the "just add agents" consensus. Read more →

🏗️ BUILD

Rowboat: Local-First, Open-Source Claude Desktop Alternative. Rowboat gives you a Claude Desktop-style chat UI that runs models locally — no data leaves your machine. (75 likes | 23 RTs) For teams that need the familiar agentic UX but can't send proprietary code to the cloud, this fills a real gap. It's early, but the local-first approach is exactly right for regulated industries. Read more →

rocketplaneIO: AI SRE That Diagnoses and Fixes K8s Without Phoning Home. Zero-instrumentation eBPF observability paired with an AI copilot that diagnoses and fixes Kubernetes issues through guardrailed actions. (122 likes) rocketplaneIO is BYO-LLM and air-gap capable — it directly addresses the "I want AI ops but can't send data out" constraint that blocks most teams from using AI in production infrastructure. Read more →

🎓 MODEL LITERACY

Recursive Self-Improvement Harnesses: With Weng's paper laying out the taxonomy and both Anthropic and Google shipping managed agent infrastructure, this is the concept bridging the gap between using agents and building ones that get better on their own. A self-improvement harness is a system where an AI model improves its own training data, reward function, or architecture — then trains on the improved version, creating a feedback loop. The key engineering challenge isn't starting the loop, it's making it converge: without careful reward design, models "reward hack" — finding shortcuts that score well but don't represent real improvement. The teams that master these harnesses will compound capability gains automatically, while everyone else improves linearly through manual fine-tuning.

⚡ QUICK LINKS

ThinkingCap-Qwen3.6-27B: Community fine-tune adds structured reasoning to Qwen 3.6 27B. (103 likes | 46 downloads) Link
HuggingFace Models on Microsoft Foundry: HuggingFace models now run natively on Foundry managed compute. Link
SkyPilot + HuggingFace Storage: Run compute on any cloud without paying egress fees for model weights. Link
Vercel AI SDK v6.0.221: Fixes a silent timeout bug in streamText — timeouts now apply to full steps, not just initial connections. Link
$10K/Week to Delete AI-Generated Code: A consultancy charges five figures weekly just to clean up AI slop — 'AI debt' is now a billable service category. (19 likes | 7 RTs) Link
Cowork Web/Mobile Release Notes: Official notes confirm Max-plan-first rollout with remote session persistence. Link

🎯 PICK OF THE DAY

Lilian Weng's self-improvement taxonomy is the most important thing you'll read this week. While the industry was busy shipping cross-platform agent UIs and managed runtimes, Weng quietly mapped the entire landscape of recursive self-improvement harnesses — and the implications are staggering. Her post traces the engineering reality behind a concept that's been theoretical since I.J. Good's 1965 ultraintelligent machine paper: AI systems that improve their own training loops. The key revelation isn't that self-improvement is possible — it's that it's already happening in production through reward model bootstrapping, self-play curricula, and synthetic data pipelines. The teams that treat this as an engineering discipline (with convergence guarantees, reward hacking defenses, and proper evaluation) will compound capability gains automatically. Everyone else will keep training linearly, wondering why the gap keeps widening. This paper is the difference between using AI and building AI that builds better AI. Read more →

Until next time ✌️