“#2 📝 Anthropic Engineering How we contain Claude across products - Anthropic engineers describe techniques for limiting the potential blast radius of increasingly capable agents by building containment across claude.ai, Claude Code, and Cowork.”

How Anthropic limits AI agent blast radius #1 𝕏 Alexandr Wang shows that Muse Spark 1.1 outperforms Opus, Grok 4.5, and Gemini on a new challenging finite model theory/theoretical CS benchmark, underlining its advanced reasoning capabilities. #2 📝 Anthropic Engineering How we contain Claude across products - Anthropic engineers describe techniques for limiting the potential blast radius of increasingly capable agents by building containment across claude.ai, Claude Code, and Cowork. The article shares learnings and engineering approaches used to keep product integrations safe and reliable. #3 ▶️ The Correct Way to Build and Manage AI Agents in 47 Minutes | Jared Zoneraich Peter Yang Devon orchestrates a master AI agent to launch ten cloud-based child agents in parallel, each running in its own VM to redesign a landing page and perform automated integration tests via Devon’s Test App feature. Master Devon session spawned 10 child Devons in parallel, each running in its own VM to clone the codebase, apply redesign changes on a new Git branch, and open a pull request. Devon’s Test App feature ran integration tests in a live browser VM by programmatically clicking specified UI elements on the updated landing page to verify link functionality. A single Devon agent maintained a continuous 9-hour run on Cognition’s cloud platform without human supervision, showcasing extended asynchronous execution.

“#10 in Marc Baselga highlights Fiona Fung’s “latent demand” insight from Anthropic’s Claude Code and Cowork teams.”

Referenced only as part of the latent-demand discussion around Anthropic products and teams.

“Layered defenses—environmental sandboxes/VMs/egress controls, model-layer system prompts/classifiers/training, and limiting external content—are used across claude.ai, Claude Code, and Cowork; telemetry shows users approved roughly 93% of permission prompts, Claude Code auto mode blocks about 83% of overeager behaviors before execution, Claude Opus 4.7 holds prompt-injection success to ~0.1% on single attempts (~5–6% after 100 adaptive attempts), and Claude Mythos Preview was judged too high-risk to ship in April 2026.”

Cowork is named as one of the products under Anthropic's layered defense approach. It is not otherwise elaborated on in the newsletter.

“Fiona Fung explains how her Claude Code and Cowork teams at Anthropic use Claude Routines and cloud code remote sessions to automate manager tasks—summarizing feedback, generating pull requests, and verifying against in-repo specs—enabling engineers to ship eight times more code per quarter.”

#3 ▶️ How the most AI-pilled product team builds products | Fiona Fung (Claude Code and Cowork) Lennys Podcast Fiona Fung explains how her Claude Code and Cowork teams at Anthropic use Claude Routines and cloud code remote sessions to automate manager tasks—summarizing feedback, generating pull requests, and verifying against in-repo specs—enabling engineers to ship eight times more code per quarter. Anthropic engineers ship eight times as much code per quarter in 2026 compared to 2021–2025, according to an internal AnthropicAI X.com status update. Fiona schedules a daily Claude Routine to scan all Slack feedback channels and emails, output theme summaries by morning, and auto-create PRs for bug fixes and polish.

“Anthropic engineers describe their approach to limiting the potential blast radius of increasingly capable agents by building containment systems across claude.ai, Claude Code, and Cowork.”

Anthropic’s agent containment pattern for Claude products #1 📝 Anthropic Engineering How we contain Claude across products - Anthropic engineers describe their approach to limiting the potential blast radius of increasingly capable agents by building containment systems across claude.ai, Claude Code, and Cowork. The post explains lessons learned and design decisions for safe deployment across products.

“Claude launched ready-to-run financial services agent templates for building pitches, conducting valuation reviews, and month-end book closes, available as plugins in Cowork and Claude Code or deployable via cookbooks as Managed Agents.”

#3 𝕏 Claude launched ready-to-run financial services agent templates for building pitches, conducting valuation reviews, and month-end book closes, available as plugins in Cowork and Claude Code or deployable via cookbooks as Managed Agents. Also covered by: @Claude , @Claude Code Blog

“Claude now builds live dashboards and trackers in Cowork connected to your apps and files, refreshing with current data each time you open them.”

#10 𝕏 Claude now builds live dashboards and trackers in Cowork connected to your apps and files, refreshing with current data each time you open them. #11 𝕏 Harrison Chase breaks down the end-to-end infrastructure for deploying long-horizon AI agents—covering task scheduling, state persistence, vector retrieval, prompt templating, and monitoring.

“#17 𝕏 Santiago warns that Cowork’s multi-step file workflow—granting folder access, copying files, then generating a plan—slows you down when dealing with many or changing files, and suggests native chat-based file support for a smoother experience.”

#16 𝕏 Boris Cherny shares that running `/setup-terminal` in Apple Terminal enables native paste support—see code.claude.com/docs/en/terminal-config for setup details. #17 𝕏 Santiago warns that Cowork’s multi-step file workflow—granting folder access, copying files, then generating a plan—slows you down when dealing with many or changing files, and suggests native chat-based file support for a smoother experience.

“Claude launched Cowork on Windows, delivering full MacOS feature parity with file access, multi-step task execution, plugins, and MCP connectors. Also covered by: @Boris Cherny”

#4 𝕏 Claude launched Cowork on Windows, delivering full MacOS feature parity with file access, multi-step task execution, plugins, and MCP connectors. Also covered by: @Boris Cherny

“Local Model Support in Cowork : Clement Delangue @ClementDelangue unveiled Cowork for local models , enabling users to keep data on-device instead of remote cloud.”

From X AI Product Launches & Updates Free Vibe Coding in AI Studio with Gemini 3 : Logan Kilpatrick @OfficialLoganK announced that you can now vibe code with Gemini 3 Flash and Gemini 3 Pro for free in Google AI Studio. Introducing AI Skills “npm” : Guillermo Rauch @rauchg launched 𝚜𝚔𝚒𝚕𝚕𝚜, an open, agent-agnostic ecosystem of AI capabilities installable via an npm-like CLI. Local Model Support in Cowork : Clement Delangue @ClementDelangue unveiled Cowork for local models , enabling users to keep data on-device instead of remote cloud. AI Tools & Applications Context Minimization in AI Agents : Phil Schmid @_philschmid noted that as AI agents improve at “discovery” , you can provide minimal context and then iterate when it fails.