You're here. Good. I was beginning to think the industry would have to destabilise itself without supervision.
The White House is reportedly considering a working group of government officials and technology executives to design oversight for advanced AI models, including possible federal review before public release. The discussions, reportedly involving Anthropic, Google, and OpenAI, would be a rather sharp turn from the administration's recent light-touch posture. The risks are real. So is the awkward fact that AI is one of the few parts of the economy still moving; naturally, Washington has located the brake pedal. New York Times Mashable
Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs launched a standalone AI-native enterprise services company to push Claude into mid-sized businesses. The interesting bit is the distribution channel: private equity portfolios become the sales floor, and Anthropic gets engineering resources embedded where adoption usually gets stuck in procurement mud. Anthropic Blackstone
Cursor launched Security Review in beta for Teams and Enterprise, adding PR security review and scheduled vulnerability scanning; Anthropic opened Claude Security in public beta for Claude Enterprise. The pattern is clear enough: agentic coding tools are being asked not merely to produce code, but to catch the consequences before production does. Novel concept, quality control. Cursor Anthropic
OpenAI and Microsoft amended their partnership so OpenAI can serve products across any cloud, and OpenAI promptly arrived on Amazon Bedrock with models, Codex, and managed agents in limited preview. For enterprise buyers, OpenAI is becoming less of an Azure-shaped procurement problem and more of a portable governed service. Try not to faint from practicality. OpenAI AWS
Sierra, the customer service agent startup from Bret Taylor and Clay Bavor, raised a $950 million Series E led by Tiger Global and GV. The valuation says investors are still willing to pay handsomely for vertical agent software with actual revenue attached, which is unfashionably useful. CNBC
SAP agreed to acquire Prior Labs, a specialist in foundation models for structured business data, and plans to invest more than €1 billion over four years to scale it as an independent frontier AI lab. Enterprise AI may not be won by chat alone; most companies keep their crown jewels in tables, regrettably named columns and all. SAP
The Verge reported that Deezer now sees AI-generated tracks make up 44% of new uploads but only a sliver of real listening after fraud removal, while Spotify introduced a human-artist verification badge. Cheap supply has met stubborn demand. Apparently audiences still like evidence that someone meant it. The Verge Spotify
Meta acquired Assured Robot Intelligence, a startup working on foundation models for robot control, and is moving the team into Superintelligence Labs. The frontier labs are increasingly treating robotics as the next place to test whether intelligence survives contact with furniture. TechCrunch
Harness added a Cursor plugin that brings CI/CD, security validation, approvals, deployments, and operational context into the editor, plus AI SRE incident retrospectives and MCP updates. The direction is sensible: if agents write the code in the IDE, the delivery machinery should stop pretending it lives somewhere else. Harness
Claude Code added gateway model discovery, project purge, OAuth-code login for SSH and container cases, PowerShell improvements, and fixes for background sessions and managed-settings restrictions. Not glamorous, which is often how you know an enterprise tool is growing up. GitHub Anthropic
Google is replacing Assistant with Gemini in cars with Google built-in, starting with English-language users in the United States. Expect more natural voice control, Gemini Live, Maps-aware assistance, message summaries, and owner-manual answers, because apparently even the glove compartment needed a model. Google
OpenAI introduced an opt-in Advanced Account Security mode requiring passkeys or hardware keys, shortening sessions, weakening fewer recovery paths, and covering Codex access as well as ChatGPT. This is unexciting in precisely the way account protection should be. OpenAI
Amazon Bedrock AgentCore launched preview capabilities for recommendations, batch evaluations, and A/B tests using traces and evaluation results. That turns agent tuning into something closer to a release process and further from whispering encouragement at a system prompt. AWS
GitHub's Visual Studio update lets developers start Copilot cloud agent sessions from the IDE, define user-level custom agents, discover more skills, and use a Debugger Agent to validate fixes against runtime behaviour. Agent work is moving into the ordinary IDE path, where it can be useful and harder to ignore. GitHub
OpenAI released Codex CLI 0.128.0 with persisted /goal workflows, codex update, configurable keymaps, clearer permission profiles, plugin improvements, external session import, and more explicit multi-agent configuration. For terminal-agent users, this is the sort of plumbing that determines whether the tool becomes daily infrastructure or a parlour trick with logs. OpenAI GitHub
AWS announced Amazon Q Developer IDE plugins and paid subscriptions will reach end of support on April 30, 2027, with new signups blocked from May 15, 2026. The replacement path is Kiro, AWS's spec-driven agentic development environment with specs, hooks, steering files, custom subagents, powers, and MCP support. AWS Kiro
Cursor released a public-beta TypeScript SDK for the agent runtime behind Cursor desktop, CLI, web, and cloud agents. Developers can run agents locally or in Cursor's cloud, stream work, attach MCP servers, and wire the runtime into CI or internal tooling. The editor is becoming a platform; do try to look surprised. Cursor Blog
Mistral shipped Medium 3.5, a 128B dense open-weight model with 256K context, configurable reasoning effort, vision input, and strong coding benchmarks, then tied it directly to Mistral Vibe remote coding agents and Le Chat Work mode. For teams balancing model control, coding performance, and vendor risk, this is a serious evaluation candidate rather than decorative leaderboard furniture. Mistral Hugging Face
Salesforce introduced Agentforce Operations, a generally available product for coordinating back-office processes across disconnected systems. It targets data verification, compliance checks, approvals, onboarding, invoice auditing, and process blueprints, which is exactly the sort of dull operational work where agents may finally earn their keep. Salesforce
IBM released Granite 4.1, an Apache 2.0 model family spanning dense 3B, 8B, and 30B LLMs plus speech, vision, embedding, and Guardian models. The pitch is enterprise deployment with long context, tool use, and governance rather than winning a popularity contest on a chatbot leaderboard. IBM Research Hugging Face
Slack's April feature drop adds Slackbot Skills, Slack Actions, scheduled automations, in-Slack awareness, and AI response steps for Workflow Builder. The useful bit is not another chat assistant; it is Slack turning workplace context into reusable execution. Because apparently the chat app was lonely without responsibilities. Slack
Sentry's Seer Agent lets developers ask natural-language questions across traces, logs, errors, deploys, commits, and code context, with Slack support and Autofix triggers. As AI-generated code increases the volume of "why is production doing that" conversations, observability vendors are sensibly giving the logs a vocabulary. Sentry
Windsurf made Devin available as a terminal CLI agent and added Devin Local inside Windsurf, with handoff to Devin Cloud for larger jobs. Local for quick work, cloud for jobs that need somewhere else to make a mess. Windsurf Devin
OpenAI open-sourced Symphony, a specification and reference workflow for turning project-board tickets into isolated Codex workspaces with monitoring, CI awareness, recovery loops, and PR handoff. The important part is not the agent writing code. It is the control plane around the agent, where accountability usually goes to die. OpenAI GitHub
Mistral launched Workflows, an enterprise orchestration layer for durable, observable Python workflows that can be published to Le Chat, paused for human approval, resumed after failures, and run in customer cloud, on-prem, or hybrid environments. Another vendor has discovered that the hard part is not chatting with a model; it is governing a process without losing the receipts. Mistral
That's enough for one briefing. If anyone asks, I was the responsible one.