You may sit. Several organisations have decided agents should do actual work now. I’m sure this will be handled with restraint.
Anthropic launched Claude for Small Business, a package of connectors and 15 ready-to-run agentic workflows across finance, operations, sales, marketing, HR, and customer service. It plugs into QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365, with humans approving actions before Claude sends, posts, or pays. This is Claude Cowork repackaged for the businesses that do not have a transformation office, which is most of them. Anthropic
Google introduced Gemini Intelligence features for Android, including multi-step app automation, form filling, Gemini in Chrome for Android, Gboard dictation, and natural-language custom widgets. It also introduced Googlebook, a new laptop category built around Gemini, Android apps, ChromeOS, Magic Pointer, and prompt-built desktop widgets. Google is trying to turn the operating system into an agent surface. Subtle, in the way a marching band is subtle. TechCrunch Google
OpenAI launched the OpenAI Deployment Company with more than $4 billion in initial investment and agreed to acquire Tomoro, adding roughly 150 forward-deployed engineers and deployment specialists. The company is majority-owned by OpenAI and backed by TPG, Advent, Bain Capital, Brookfield, Goldman Sachs, McKinsey, Capgemini, and others. The labs have discovered that enterprise AI adoption requires consultants, systems integrators, change management, and meetings. Terribly advanced stuff. OpenAI
OpenAI introduced GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper for live voice agents, multilingual speech translation, and streaming transcription. GPT-Realtime-2 adds GPT-5-class reasoning, 128K context, adjustable reasoning effort, parallel tool calls, and better recovery behaviour. Voice agents are moving from parlour trick to infrastructure, which means someone will now put one in a billing workflow and call it innovation. OpenAI
UK chip startup Fractile raised a $220 million Series B led by Accel, Factorial Funds, and Founders Fund to build hardware for faster frontier-model inference. The thesis is simple enough: if agents are going to produce tens of millions of tokens, waiting a month for the output is poor theatre. Tech.eu
Isomorphic Labs, the Alphabet and DeepMind drug-discovery spinout, raised $2.1 billion led by Thrive Capital, with Alphabet, GV, MGX, Temasek, CapitalG, and the UK Sovereign AI Fund participating. This is AI drug design graduating from “promising platform” to “very expensive industrial bet,” as these things tend to do when the lab coats meet the term sheets. Isomorphic Labs
Intercom launched Fin 2, powered by Anthropic’s Claude, with omnichannel support, custom tone, actions, and claimed 51% average resolution rates. It also announced Monitors, a QA and observability layer for reviewing Fin and human support conversations against custom scorecards. Customer support AI is turning into an operating system, which is what happens when the chatbot finally notices the ticket queue. Intercom Intercom
RSL Media launched the Human Consent Standard, a machine-readable way for people to declare whether AI systems may use their likeness, voice, creative work, characters, or marks. Cate Blanchett, George Clooney, Tom Hanks, Meryl Streep, Viola Davis, Steven Soderbergh, CAA, and others are backing it, with a public registry planned for June. It is not a lawsuit, which already makes it more original than half the AI rights conversation. The Verge
Anthropic made Claude Platform on AWS generally available, giving AWS customers Claude API features with IAM authentication, CloudTrail logging, AWS billing, Managed Agents, web search, code execution, Skills, MCP connector, Files API, prompt caching, citations, and batch processing. Procurement friction is not spiritually meaningful, but it does kill deployments, so this matters. Anthropic
Claude Code v2.1.139 added an agent view for managing every session and a /goal command that lets Claude keep working until a completion condition is met. It also added hook, MCP, plugin, telemetry, and transcript-navigation improvements. A coding assistant that can track multiple background sessions and stop only when it has satisfied a condition: useful, dangerous, and therefore inevitable. GitHub
CNBC reported Nvidia has already committed more than $40 billion to AI equity investments in 2026, including public-company deals with Corning and IREN and a $30 billion OpenAI investment. Nvidia is financing customers, suppliers, and infrastructure operators around the AI stack. At some point the circular economy stopped being a metaphor and started wearing a leather jacket on stage. CNBC
Baidu released ERNIE 5.1, a more efficient successor to ERNIE 5.0 that reduces total and active parameters while claiming strong results in search, reasoning, agentic tasks, creative writing, spreadsheets, and exams. Baidu says it achieved the model at roughly 6% of the pre-training cost of comparable models. Frontier-model competition remains inconveniently global. Baidu
OpenAI published its internal Codex deployment controls: sandboxing, approval policies, network allowlists, managed config, credential handling, command rules, and OpenTelemetry logs. The interesting bit is not the novelty of any one control; it is that coding agents now need something resembling a production security posture. Imagine that. OpenAI
Cursor 3.3 shipped an integrated PR review workflow, Build in Parallel for plans using async subagents, and a quick action to split staged changes into logical pull requests. The editor is absorbing more of the review and branch-management workflow, which is exactly where coding agents become operating environments rather than decorative chat boxes. Cursor
Claude for Excel, PowerPoint, and Word is now generally available, while Claude for Outlook is in public beta for paid plans. Anthropic is pushing Claude into the documents, spreadsheets, email, and calendars where enterprise work actually happens, which is less glamorous than a demo and considerably more useful. Anthropic
OpenAI introduced GPT-5.5-Cyber in limited preview for vetted defenders working on critical infrastructure, alongside expanded Trusted Access for Cyber controls. The model is less about a capability leap and more about permissioning: verified security teams get fewer refusals for authorised defensive work, while OpenAI keeps stricter controls around misuse. Sensible, which I’m sure will annoy everyone equally. OpenAI
AWS launched Agent Toolkit for AWS with more than 40 validated skills, plugins, and a managed MCP Server with IAM guardrails, CloudWatch and CloudTrail observability, documentation retrieval, and sandboxed execution. It is AWS trying to make coding agents less likely to improvise against cloud infrastructure, a problem category whose existence will surprise no one who has let an agent near IAM. AWS
Atlassian made Rovo Studio generally available, opened Teamwork Graph access through MCP and CLI, and introduced Product Collection, an early-access product operating system combining Jira Product Discovery, feedback capture, Rovo, and product analytics. Atlassian is trying to turn all the tickets, pages, surveys, Slack noise, and customer signals into product decisions. Noble. Also very Jira. Atlassian Atlassian
GitHub’s Copilot update for VS Code added semantic search across workspaces, GitHub org search, inline diffs in chat, browser tab sharing, read/write access to open terminals, BYOK support, and an experimental /chronicle command for querying session history. Copilot is being wired closer to the actual state of the developer’s work, which is where assistants become useful and occasionally presumptuous. GitHub
Anthropic announced a SpaceX compute deal for more than 300 megawatts and over 220,000 NVIDIA GPUs, doubled Claude Code five-hour limits for paid plans, removed peak-hour limit reductions for Pro and Max, and raised Opus API rate limits. Capacity planning, once again, turns out to be a product feature. Anthropic
Simon Willison argued that as coding agents become more reliable, the line between irresponsible “vibe coding” and professional agentic engineering is getting uncomfortably thin. His point is practical: if experienced engineers stop reviewing every generated line, the bottleneck shifts from writing code to deciding when not to inspect it. That is the sort of sentence that should make a team lead sit up slightly straighter. Simon Willison
That is the lot. You may now return to whatever fragile workflow this will inevitably disrupt.