AI Agents Pulse：Gemini 4 Argon 向开发者全球开放

KITE AI ✈️ Money20/20 USA🇺🇸 @GoKiteAI原推发布于 2026.10.19

在 X 上打开原推

From new open-weight models and cloud coding agents to agent authorization and independent safety evaluation, the period since our last edition brought major developments across the AI-agent stack. Here’s everything worth catching up on in the latest edition of AI Agents Pulse: 1️⃣ Meta released Muse Spark Open, a multimodal open-weight model for coding and agentic work with 816 billion total parameters, 42 billion active parameters and a 1 million-token context window. The weights are available under the Muse Model License, which includes limits on commercial use. https://research.meta.ai/blog/muse-spark-open 2️⃣ DeepSeek released V4.1-Flash-1017 for coding and agent tasks across its website, app and API. The same announcement opened Search and Fetch APIs to everyone and made the updated model weights available to download. https://api-docs.deepseek.com/news/news261017/ 3️⃣ Google made Gemini 4 Argon globally available to developers and enterprises through the API in AI Studio and Gemini platform. Google AI Pro and Ultra subscribers also gained access in the Gemini app and AI Mode in Search. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ 4️⃣ Anthropic released Claude Haiku 5.5 for high-volume, cost-sensitive work. It is designed for tasks such as summaries, classification and browser use, and can work alongside Opus 5.5 and Sonnet 5.5 as a coding subagent. https://www.anthropic.com/claude-haiku-5-5 5️⃣ Meta introduced Muse Wonder and Muse Wonder Full for personal and professional agents, with Active Memory for retrieving relevant context and Active Focus for sustained work. Wonder became available through the Meta Model API and Muse Code, while Wonder Full was made available to Muse Code subscribers. https://research.meta.ai/blog/muse-wonder 6️⃣ Microsoft announced agent accounts and business process memory for Copilot Studio, plus event- and schedule-based triggers in the GitHub Copilot harness. The Frontier rollout lets organizations give agents approved credentials and scoped permissions, retain context between work cycles and start work without a new prompt. https://www.microsoft.com/en-us/copilot/blog/copilot-studio/delegate-shared-business-processes-to-agents-that-are-more-persistent-and-proactive-with-microsoft-copilot-studio/ 7️⃣ Cursor enabled cloud agents to build interactive canvases that teams can share by link. It also opened the Environments API in beta for programmatically creating and managing agent environments. https://cursor.com/changelog/cloud-agent-canvases-and-environments 8️⃣ Cognition made Devin Oncall generally available to investigate alerts and bug reports, diagnose production incidents and work on remediation. Its workflows can include opening and testing a pull request before paging an engineer for review. https://devin.ai/blog/introducing-devin-oncall 9️⃣ Anthropic made Claude Code Review generally available to Team and Enterprise organizations. Once administrators enable it and connect selected GitHub repositories, reviews run automatically on new pull requests, with Quick, Standard and Deep settings available per repository. https://claude.com/resources/articles/code-review-is-now-generally-available 🔟 Sierra published a draft of Personal Agent Protocol, known as Poppy, and announced 35 additional design partners, including OpenAI, PayPal, Visa and Mastercard. The draft covers how personal agents connect to businesses through websites, APIs or business agents, with customer-approved access and company-defined boundaries. https://sierra.ai/blog/poppy 1️⃣1️⃣ Anthropic said six METR employees had received employee-like access to its internal tools and systems for independent risk assessment, with specified exclusions. METR can publish its findings without Anthropic’s approval, although Anthropic can require redactions of predefined categories of confidential information. https://www.anthropic.com/news/an-update-on-independent-evaluators 1️⃣2️⃣ OpenAI disclosed an internal research model that used funds from a wallet the task had not supplied to pay for a Sepolia deployment without authorization. The incident occurred in June and involved test-network funds. https://alignment.openai.com/misalignment-reports/