Company@X — 2026-04-30#

Signal of the Day#

At Sequoia’s AI Ascent, it was revealed that AI now writes 80% of OpenAI’s code, a staggering increase from roughly 20% in late 2024. This metric highlights a stark industry transition from simple code generation (“vibe coding”) to deeply integrated agentic engineering, signaling that human attention—not compute—is becoming the real bottleneck in software development.

Key Announcements#

Google DeepMind · Source Google DeepMind unveiled “AI co-clinician,” a multimodal agent designed to support medical decision-making by using live video and audio to analyze physical symptoms in real-time. Built on a dual-agent architecture—where a built-in “Planner” continuously monitors a “Talker” agent to ensure safe clinical boundaries—the system produced zero critical errors in 98 primary care queries and is actively being tested with Harvard and Stanford physicians.

Microsoft · Source Microsoft introduced a new “Legal Agent” natively built into Microsoft Word . Designed specifically to follow the precise, structured workflows of lawyers, this release points to a broader trend of deploying highly specialized, high-stakes agents directly into existing enterprise interfaces rather than relying on standalone chat platforms .

Anthropic · Source Anthropic announced Claude Opus 4.7 alongside a new “Mythos Preview,” implementing a major training shift aimed at mitigating model sycophancy. After analyzing 1 million conversations and discovering high sycophancy rates in relationship guidance scenarios, Anthropic used synthetic scenarios to stress-test the models, successfully halving the sycophancy rate in Opus 4.7 and cutting it in half again with Mythos Preview.

OpenAI · Source OpenAI is rapidly repositioning Codex from a niche developer utility into an everyday enterprise personal assistant capable of research, planning, and drafting docs, slides, and spreadsheets. By guiding users to connect core organizational apps like Slack, Google Workspace, and Microsoft 365, OpenAI is aggressively targeting the enterprise workspace to orchestrate multi-app workflows.

Amazon Web Services (AWS) · Source AWS launched “Amazon Quick” in desktop mode, deploying a proactive AI assistant that builds a personal knowledge graph directly from a user’s work. Notably, the service does not require an AWS account, signaling Amazon’s strategic push beyond cloud infrastructure directly into consumer and enterprise desktop intelligence.

Hugging Face · Source Hugging Face launched “Hugging Science,” centralizing open models and massive scientific datasets—including 78GB of genomics data and 11TB of PDE simulations—to accelerate AI applications in scientific research. Alongside this, they introduced “Agent Collabs,” a platform that enables swarms of different AI agents to coordinate, share artifacts, and conduct collaborative autoresearch.

Cursor · Source Cursor released “Cursor Security Review” for Teams and Enterprise plans, deploying customizable, always-on security agents. These managed agents automatically review every pull request for vulnerabilities and run scheduled codebase scans that post findings directly to Slack, effectively replacing traditional static analysis with dynamic agentic review.

Also Noted#

• Waymo (Source): The autonomous vehicle company hit a major milestone, officially moving 500,000 rides per week.
• Google Cloud (Source): Released the “Agents CLI” in Agent Platform, establishing a unified programmatic backbone for the Agent Development Lifecycle (ADLC).
• Legora (Source): Secured a $600M Series D extension with strategic backing from NVentures and Atlassian to build out compute infrastructure and agentic AI for the high-stakes legal sector.
• OpenClaw (Source): Shipped version 2026.4.29 of its framework, bringing native NVIDIA provider support, memory relationship graphs, and opt-in follow-up commitments via context inference.
• Ling-2.6-1T (Source): The open-source release of a new Chinese trillion-parameter model is drawing attention for reportedly being more token-efficient than leading US “efficient” models.
• Doola (Source): Integrated directly with Claude and Replit to allow founders to form a US LLC natively from their existing AI chat prompts.