Sit down. The industry has once again confused velocity with navigation.
Adobe launched a productivity agent for Acrobat that can chat across PDFs, extract insights, generate presentations, podcasts, blogs, and social posts, and package material into interactive PDF Spaces. It is a rather direct attempt to make the PDF less of a static artefact and more of a governed work surface, which is useful if your company still conducts half its business through attachments named final_v7_really.pdf. Adobe Business Wire
Mistral released Medium 3.5, a 128B dense open-weights model with 256K context, vision, reasoning controls, and strong coding benchmarks, then made it the default model for Le Chat and Mistral Vibe remote coding agents. For teams wary of handing the entire development loop to one American hyperscaler, this is a credible evaluation candidate rather than decorative leaderboard furniture. Mistral
Anthropic launched ten financial-services agent templates for pitchbooks, KYC screening, valuation review, month-end close, and related work, available through Claude Cowork, Claude Code plugins, and Managed Agent cookbooks. Claude add-ins for Excel, PowerPoint, and Word are also generally available, which means Claude is now being invited directly into the spreadsheet habitat. Do keep your macros calm. Anthropic
The Trump administration is reportedly considering a government-and-industry working group that could design federal oversight for advanced AI models, including possible pre-release review. The risks are not imaginary, but neither is the economic dependence on these systems; naturally, Washington has found the brake pedal just as the engine began making interesting noises. New York Times Mashable
Willison argues that as coding agents become more capable, even experienced engineers are beginning to treat agent output as semi-black-box work rather than reviewing every line. That should make everyone slightly uncomfortable, which is usually where useful engineering discipline begins. Simon Willison
The UK government announced legislation requiring platforms and AI chatbots to remove non-consensual intimate images, including deepfakes, within 48 hours of a report. This is regulation aimed at a real measured harm, not merely a committee being startled by a chatbot. Computing
OpenAI opened a beta self-serve Ads Manager for U.S. advertisers, adding cost-per-click bidding, conversion measurement, and pixel-based attribution while saying ad delivery remains separate from ChatGPT answers. The assistant is acquiring a sales department. I’m sure this will be handled with exquisite restraint. OpenAI OpenAI
Google expanded Gemini API File Search so developers can build retrieval apps over images and text together, with metadata filters and page-level citations. Multimodal RAG with citations is precisely the sort of unglamorous plumbing that determines whether an enterprise assistant is useful or merely confident near a document. Google
GitHub made secret scanning through its MCP Server generally available and added dependency vulnerability scanning in public preview. AI coding agents can now check current changes for exposed secrets and vulnerable packages before commit or PR, which is better than discovering the secret after it has joined the public internet and developed a social life. GitHub GitHub
Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs launched a standalone AI services company to bring Claude into mid-sized companies. The distribution channel is the interesting part: private equity portfolios become the deployment surface, and Anthropic gets engineering help where adoption usually dies in procurement. Anthropic Blackstone
Sierra raised $950 million led by Tiger Global and GV, saying it now serves more than 40% of the Fortune 50 across customer-service-heavy industries. Investors are still willing to pay handsomely for vertical agent software with actual revenue attached, which is unfashionably practical. Sierra TechCrunch
SAP agreed to acquire Prior Labs and invest more than €1 billion over four years in a frontier AI lab for structured business data. Enterprise AI may not be won by chat alone; most companies keep their crown jewels in tables, regrettably named columns and all. SAP TechCrunch
Andon Labs described Mona, an AI agent managing a real cafe in Stockholm, handling supplier orders, hiring, permit applications, and Slack coordination with human baristas. The result included functioning operations and obvious failures, which is more honest than most agent demos and therefore considerably more useful. Andon Labs Simon Willison
Zendesk’s advanced email AI agents can now understand messages, answer questions, automate procedures, and escalate when needed. Support is one of the places agents can deliver value without pretending to be a strategic visionary, which is refreshing. Zendesk Zendesk
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. Agentic coding tools are now being asked not merely to produce code, but to catch the consequences before production does. Novel concept, quality control. Cursor Anthropic
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 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
Cursor released @cursor/sdk, a public-beta TypeScript SDK for running Cursor agents locally or in Cursor cloud VMs, with durable agents, prompt runs, streaming, reconnect support, lifecycle controls, and standardised responses. The editor is becoming a platform; do try to look surprised. Cursor Cursor Docs
IBM released Granite 4.1, an Apache 2.0 model family spanning dense 3B, 8B, and 30B language models plus updated speech, vision, embedding, and Guardian safety 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
That is enough instability for one briefing. Use it before someone turns it into a keynote.