Sources

Company@X — 2026-03-25#

Signal of the Day#

The launch of the ARC-AGI-3 benchmark and the 2026 ARC Prize sets a definitive, rigorous standard for measuring agentic intelligence. With frontier models currently scoring under 1% while human baselines sit at 100%, the industry is drawing a hard line: true AGI requires on-the-fly world modeling and continual learning in interactive environments, not just static reasoning.

Key Announcements#

[Harvey] · Announcement Harvey announced a new funding round at an $11B valuation, led by GIC and Sequoia, aimed at scaling their long-horizon AI agents for legal teams. This signals massive enterprise confidence in the transition from AI as an advisory copilot to autonomous agents that execute complex, multi-step workflows while lawyers shift strictly to driving strategy and outcomes.

[Google DeepMind] · Post DeepMind launched Lyria 3 Pro, expanding generative audio capabilities to produce high-fidelity music tracks up to three minutes in length. Rolling out via the API in Google AI Studio and to paid Gemini subscribers, the model grants structural control over intros, verses, and choruses, positioning Google aggressively in professional creative workflows.

[Cursor] · Announcement Cursor shipped a major enterprise feature allowing their cloud agents to run directly on customer infrastructure. By keeping code and tool execution entirely within a company’s own network, Cursor is removing a critical security bottleneck for enterprise engineering teams looking to adopt agentic coding capabilities.

[Anthropic] · Engineering Blog Anthropic published engineering details on Claude Code’s “auto mode,” which utilizes internal classifiers to make execution approval decisions autonomously. This highlights the industry’s broader push to safely reduce human-in-the-loop friction in agent workflows without entirely dropping safeguards.

[Google Cloud] · Post Google Cloud introduced BigQuery global queries, a feature that enables querying data stored across different geographic locations with a single standard SQL query. By eliminating the need for manual ETL pipelines to consolidate data, this release drastically simplifies global analytics and compliance for multinational organizations.

[AWS] · Case Study AWS highlighted a major enterprise deployment with NTT DOCOMO, utilizing Amazon Bedrock AgentCore to orchestrate multiple AI agents for telecommunications network maintenance. This serves as a strong proof point for the commercial viability of orchestrated agentic AI in highly complex, mission-critical infrastructure.

Also Noted#

  • [a16z] (Announcement): Led a $5M round in Airbase to build software-defined infrastructure for modernizing radio frequency (RF) spectrum licensing and intelligence for defense and commercial sectors.
  • [Hugging Face] (Post): Developers are increasingly treating HF Buckets as a viable AWS S3 alternative for ML pipelines, utilizing hf sync to manage massive dataset and checkpoint storage seamlessly.
  • [OpenClaw] (Release Notes): Shipped version 2026.3.24, enabling sub-agent communication via Open WebUI alongside native orchestration in Slack, Teams, and Discord.
  • [Ginkgo Bioworks] (Podcast): Revealed details of an ongoing partnership with OpenAI, wherein GPT-5 was utilized to run their Autonomous Lab.
  • [Tesla] (Highlight): Elevated user reports that FSD’s environmental audio detection effectively pulled over for an ambulance, showcasing the platform’s utility as a critical accessibility layer for deaf drivers.