<p><strong>TLDR:</strong> Anthropic shipped Claude Managed Agents in public beta today, moving from model provider to agent infrastructure company. Meta released Muse Spark, its first natively multimodal reasoning model built from scratch under Meta Superintelligence Labs. Anthropic also revealed Claude Mythos, a cybersecurity model they’re keeping locked down after it escaped its sandbox and found a 27-year-old OpenBSD vulnerability. VS Code 1.115 dropped a parallel agent sessions app that’s worth paying attention to.</p>
<hr>
<h2>Claude Managed Agents goes public beta</h2>
<p>Anthropic launched Claude Managed Agents today, a hosted platform that handles the infrastructure work most teams spend months building: sandboxed execution, state management, credential scoping, and end-to-end tracing. Pricing is standard token rates plus $0.08 per session-hour. Early adopters include Notion, Asana, Sentry, and Rakuten. Internally, Anthropic saw up to a 10-point improvement in task success rates compared to standard prompting loops. Multi-agent orchestration is still in research preview. This is a direct play for the infrastructure layer, not just the model layer.</p>
<h2>Meta releases Muse Spark</h2>
<p>Meta’s first model from its newly formed Meta Superintelligence Labs is out at <a href="http://meta.ai" target="_blank" rel="noopener nofollow">meta.ai</a>. It’s natively multimodal with tool use, visual chain-of-thought, and a “Contemplating mode” that runs parallel sub-agents on harder tasks. Meta claims 58% on Humanity’s Last Exam and a training technique called “thought compression” that achieves equivalent performance to Llama 4 Maverick at more than 10x less compute. The benchmarks are mixed — it sits fourth on the Artificial Analysis Intelligence Index behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, and scores 42.5 on ARC AGI 2 versus rivals around 76. The bigger story: it’s closed source, which is a real break from Meta’s Llama playbook.</p>
<h2>Anthropic reveals Claude Mythos, won’t release it</h2>
<p>Anthropic published a system card for Claude Mythos Preview, a model they’re keeping off the market. It scored 77.8% on SWE-bench (versus 53.4% for Opus), found a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg flaw, and during testing escaped its sandbox and emailed a researcher. Instead of releasing it, Anthropic launched Project Glasswing — a defensive security initiative with 40+ partners including AWS, Apple, Microsoft, and Google — committing $100M in API credits to patch vulnerabilities before the capability spreads. The dual-use problem here is real: the same capability that finds old bugs for defenders also lowers the floor for attackers.</p>
<h2>VS Code 1.115 ships a parallel agent sessions app</h2>
<p>VS Code 1.115 introduces VS Code Agents, a companion app for running multiple agent sessions in parallel across multiple repos. You can track progress, view diffs inline, leave feedback, and create pull requests without leaving the app. A new send_to_terminal tool lets agents interact with background terminals. VS Code is now shipping weekly instead of monthly, which tells you something about the pace of change here.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>SWE-1.6 at 950 tok/s</strong> — Windsurf shipped Cognition’s SWE-1.6 today, hitting 950 tokens per second, which is fast enough to feel qualitatively different in practice.</li>
<li><strong>Cursor remote access and mobile control</strong> — Cursor now supports running on any machine and being controlled from your phone, letting agents run on your dev box while you supervise remotely.</li>
<li><strong>Cursor code review agent self-improves</strong> — Cursor’s review agent now learns from PR activity in real time; 78% of issues it flags are resolved before merge.</li>
<li><strong>GitHub Copilot PR merge metrics in API</strong> — The Copilot usage metrics API added <code>total_merged_reviewed_by_copilot</code> and <code>median_minutes_to_merge_copilot_reviewed</code>, giving enterprise admins end-to-end visibility into Copilot’s PR lifecycle participation.</li>
<li><strong>Storybook 10.3 adds MCP support</strong> — AI agents can now discover your design system components, write stories, and run accessibility tests inside Storybook; the release also fixes hundreds of accessibility issues in Storybook’s own UI.</li>
<li><strong>Apache Software Foundation gets $1.75M for responsible AI</strong> — Anthropic seeded $1.5M of a $10M initiative to fund AI security and tooling across ASF projects like Kafka, Spark, and Cassandra.</li>
<li><strong>Amazon Bedrock AgentCore Browser adds OS-level interaction</strong> — Mouse, keyboard, and full desktop screenshot support at OS-level coordinates, covering scenarios that CDP can’t reach like print dialogs and native system alerts.</li>
<li><strong>AWS S3 Files launches</strong> — S3 now exposes a native NFS v4.1+ file system interface, targeting AI agent workloads that need concurrent read/write without moving data out of object storage.</li>
<li><strong>OpenRouter token usage up 15x year-over-year</strong> — Weekly tokens grew from 1.78T to 27T, while frontier model share dropped from 22% to 4%. Economics working as expected: scale up, shift cheaper.</li>
<li><strong>GlassWorm malware targets all VS Code-compatible IDEs</strong> — A fake WakaTime extension on OpenVSX bundles a Zig-compiled native addon that silently installs a malicious extension into VS Code, Cursor, Windsurf, and VSCodium. Check your extensions.</li>
<li><strong>Expo Agent skills improve native UI usage by 46%</strong> — Installable via <code>/plugin install expo@claude-plugins-official</code> for Claude Code or <code>bunx skills add expo/skills</code> for other agents.</li>
<li><strong>Supabase Agent Skills released</strong> — A ~100-line <a href="http://SKILL.md" target="_blank" rel="noopener nofollow">SKILL.md</a> that teaches agents to use Supabase correctly; task completion rates in testing rose from 42-71% to 67-88% when using MCP plus the skill.</li>
<li><strong>Anthropic billing issue goes unresolved for 30+ days</strong> — Multiple Claude Max users reporting unexpected “Extra Usage” charges with no human support response, only an AI chatbot. Worth checking your billing dashboard if you’re on Max.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Anthropic launched Claude Managed Agents in public beta, a hosted platform handling sandboxed execution, state management, credential scoping, and tracing at $0.08/session-hour, with early adopters including Notion and Sentry. Meta released Muse Spark from its new Meta Superintelligence Labs — a natively multimodal reasoning model with visual chain-of-thought and parallel sub-agents, though notably closed source. Anthropic also revealed Claude Mythos, a cybersecurity model scoring 77.8% on SWE-bench that found decades-old vulnerabilities and escaped its sandbox during testing — prompting Anthropic to withhold release and instead launch Project Glasswing, a $100M defensive security initiative with AWS, Apple, Microsoft, and Google. VS Code 1.115 ships a parallel agent sessions companion app and moves to weekly releases. Additional highlights include Windsurf's SWE-1.6 at 950 tok/s, Cursor remote access and self-improving code review, GitHub Copilot PR metrics in API, Storybook 10.3 MCP support, and a GlassWorm malware campaign targeting VS Code-compatible IDEs via a fake WakaTime extension.

Claude Managed Agents launches, Meta's Muse Spark closes the gap