AI News | Field Notes by Michael Nemtsev
Pen-scratch cover: Create a hand-drawn pen scratch editorial illustration showing a figure seated at a desk facing a mirror, but the reflection shows mechanica

Copilot can finally edit your files instead of just suggesting changes, autonomous coding agents like Devin are raising hundreds of millions based on unproven safety claims, and the technical standards for building agents have stopped shifting every few weeks. Meanwhile, even the scrappy labs are taking money from the same giants (DeepSeek just opened talks with Tencent and Alibaba after two years of independence), and Elon Musk dropped his fraud case against OpenAI, removing one of the few external checks on how the company treats its nonprofit mission. The question of who reviews what an AI agent does while no one is watching has stopped being philosophical and become your actual job security.

AI AgentsAI Industry ·The Verge

Microsoft finally lets Copilot actually touch your documents

AnalysisAfter two and a half years of Copilot sitting in a sidebar and politely answering questions, Microsoft this week flipped Agent Mode to generally available in Word, Excel, and PowerPoint. The move is a tacit admission: Sumit Chauhan, corporate VP of Microsoft's Office Product Group, wrote that when Copilot first shipped, "foundation models were not powerful enough" to command the apps directly. Now they are, apparently. Agent Mode lets Copilot take multi-step actions straight on the document canvas, building pivot tables, reformatting decks, and rewriting reports without the copy-paste intermediary. A separate feature called Office Agent, running on Anthropic's Claude models, handles PowerPoint and Word creation from a chat window. Both are on by default for Microsoft 365 Copilot, Premium, and Personal and Family subscribers, starting on web versions of the apps first. The honest subtext: this is what Microsoft promised in 2023, finally arriving.

AI AgentsAI Industry ·Bloomberg

Devin's maker doubles to $25 billion

AnalysisCognition AI, the New York lab behind autonomous coding agent Devin, is in early talks to raise hundreds of millions of dollars or more at a $25 billion valuation, more than doubling its last round, Bloomberg reported on April 23. Cognition picked up Windsurf assets last year after Google's acqui-hire stripped out the leadership team, and has been pitching enterprises on agents that can write code with minimal supervision over multi-hour runs. The AI coding market is now visibly splitting into three categories: IDE-native (Cursor), terminal-native (Anthropic's Claude Code, OpenAI's Codex CLI), and full-autonomy (Devin). Each is a different bet on how much human review the work needs. Capital, for now, is subsidizing the answer. None of the autonomy claims have an independent evaluation harness anyone agrees on.

AI Industry ·Fortune

Musk folds his OpenAI fraud case

AnalysisElon Musk dropped his fraud claims against OpenAI and Sam Altman on April 25, days before the Musk v. Altman trial was scheduled to begin. The original 2024 suit alleged OpenAI had betrayed its founding nonprofit mission by converting to a capped-profit structure. The dismissal lands the same week SpaceX announced its $60 billion option on Cursor and Musk pivots toward making xAI credible on agents and coding. Some claims, including breach of contract, remain alive. Musk's legal pressure was one of the few external checks on OpenAI's nonprofit-to-for-profit conversion. Investors anchoring OpenAI's $852 billion post-money valuation now have one less obstacle. So does a possible OpenAI public listing.

AI Agents ·DEV Community

The agent stack stops moving every week

AnalysisMicrosoft shipped Agent Framework 1.0 in mid-April with stable APIs and full Model Context Protocol support (MCP, Anthropic's open standard for plugging tools into AI agents), plus a browser DevUI that visualizes agent execution. Claude Desktop and Cursor both shipped MCP v2.1, making tool discovery consistent across major clients. Google's Agent-to-Agent Protocol turned one year old on April 9 with more than 150 organizations participating, 22,000 GitHub stars, and production deployments inside Azure AI Foundry and AWS Bedrock AgentCore. The shape of the agent stack is finally settling: MCP for tool calls, A2A for agents talking to each other, AP2 for payments. For the first time, the protocols developers integrate against are not changing every six weeks.

AI Industry ·IndexBox

DeepSeek finally takes the money

AnalysisDeepSeek opened its first external financing window in mid-April 2026 and is now in talks with Tencent and Alibaba, Bloomberg reported on April 24. Tencent has proposed acquiring up to a 20% stake. The startup, owned by hedge fund High-Flyer and famous for releasing R1 in January 2025 on a self-reported $5.6 million budget, has resisted outside capital for two years. The shift comes as V4 launches on Huawei Ascend chips and as core authors of R1 and DeepSeek's earlier language model are reportedly being poached by ByteDance and Tencent. The benchmark valuation: roughly $20 billion, behind MiniMax at about $40 billion. The financing also signals a quieter truth: the indie-lab story has limits when frontier training runs cost what they cost.

Subscribe for full archive access

Every past issue, weekly deep dives, and the full back catalogue — delivered free.

Read on Substack

Want this in your inbox?

One email a day, zero hype.

A short read every morning: what actually changed in AI, and what it means for work and daily life. Free, unsubscribe anytime.