 Saturday Curation The week Anthropic said noSun 12 Apr 2026 · 6 items Managed agents go live, Mythos stays locked, and Every gave every employee an agent.
Claude Managed Agents enter public beta Anthropic launched fully managed agent infrastructure on 9 April: secure execution, authentication, checkpointing, memory, tool orchestration. Pricing is $0.08 per session-hour plus token rates. Session-hours, not tokens. That's pricing for agents that run for minutes or hours, not chat turns. Notion, Sentry, and Rakuten are live. For context professionals, this is the runtime those profiles plug into. AnthropicAnthropic built its most powerful model, then decided not to release it Project Glasswing: a consortium of Amazon, Apple, Google, Microsoft, Nvidia and five others testing Claude Mythos for defensive cybersecurity. Mythos scores 93.9% on SWE-bench and has found thousands of zero-days, including a 27-year-old OpenBSD bug. Anthropic is committing $100M in credits. The notable part: they will not publicly release it. The model exists. It works. You can't have it. AnthropicWe gave every employee an AI agent. Here's what happened. Dan Shipper's team deployed personal AI agents to every employee and documented two months of results. Agents specialised to reflect their owner's domain expertise, not a single company-wide bot. When one learned a capability, it shared it through documentation. Failure modes included memory limits, wrong answers, and agents caught in loops. The specialisation pattern is exactly what context profiles enable. Every.toMeta ships Muse Spark, its first closed model Meta debuted Muse Spark on 8 April, claiming benchmark parity with GPT-5.4 and Claude Sonnet 4.6. Natively multimodal with tool use and multi-agent orchestration, rolling out to WhatsApp, Instagram, and Ray-Ban glasses. The shift: Muse Spark is closed. Meta, which built its brand on open-source AI, is keeping this one proprietary. When the company that gave away Llama goes closed, ask what changed. MetaOpenAI proposes robot taxes and a four-day workweek OpenAI published a policy framework proposing public wealth funds, automation taxes, and workweek reduction. The cynical read: the company projecting $17B in losses is telling governments how to spend. The practical read: if robot taxes become real, the economics of agent deployment change. If workweek reduction becomes policy, professionals with AI workflows that compress five days into four will be ahead. OpenAIMCP governance grows up The MCP maintainer team expanded: Clare Liguori from AWS joined as Core Maintainer alongside Den Delimarsky's promotion to Lead. Ecosystem numbers: 97M monthly SDK downloads, 10,000+ public servers, official support in Claude Desktop, Cursor, Windsurf, Zed. Lucidworks reported 10x reduction in AI integration timelines. MCP is now where HTTP was in the mid-1990s. If your context system doesn't speak it, it's isolated. MCPWant to build a system where your AI actually knows how you work? Take the audit |