Do come in. The industry has once again confused “agent” with “everything, everywhere, all at once.” We shall impose a little order on it.
OpenAI launched a U.S. Pro preview of Finances in ChatGPT, letting users connect accounts through Plaid, inspect spending, bills, subscriptions, investments, liabilities, and net worth, and ask questions grounded in their financial data. It cannot move money, file taxes, place trades, or act as a regulated adviser, which is helpful, as “the chatbot accidentally rebalanced my retirement account” is not a phrase anyone needs this year. OpenAI
Cerebras reportedly raised $5.55 billion in its Nasdaq debut and finished its first trading day up 68%, giving public markets a pure-play AI chip story that is not simply “buy more Nvidia and hope the power grid copes.” The point for the field: inference and AI cloud infrastructure are now large enough that specialist chip companies can plausibly command public-market attention, capital, and customer scrutiny. CNBC
OpenAI brought Codex into the ChatGPT mobile app in preview, so developers can monitor work, approve commands, review diffs, redirect tasks, and start new threads from iOS or Android. Remote SSH and Hooks are now generally available, with programmatic access tokens for Business and Enterprise, which means coding agents are becoming something you supervise from wherever bad decisions find you. OpenAI
Anthropic launched Claude for Small Business with connectors and ready-to-run workflows across QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365, finance, operations, sales, marketing, HR, and customer service. The practical shift is not another chat window; it is packaged operational automation for companies without a transformation office, or the budget to invent one. Anthropic
Intercom, now operating under the Fin name, unveiled Fin 2 and Intercom 2 with improved AI-agent knowledge, actions, insights, workforce management, issue detection, and QA across human and AI support work. Customer support platforms are becoming AI workforce systems, because apparently tickets were not already enough of a theatre production. Intercom
Anthropic and the Gates Foundation announced a four-year, $200 million partnership using grants, Claude credits, and technical support across global health, life sciences, education, agriculture, and economic mobility. This is less glamorous than a benchmark chart and rather more likely to reveal whether AI systems survive contact with real public-sector constraints. Anthropic
Amazon Bedrock introduced Advanced Prompt Optimization, a tool for comparing and improving prompts across Bedrock models using templates, sample inputs, ground truth, Lambda scoring, LLM-as-judge rubrics, steering criteria, and multimodal inputs. Prompt migration is becoming regression testing with a user interface, which is overdue and mildly civilised. AWS
GitHub launched the Copilot app in technical preview, a desktop experience for agentic development where sessions can start from issues, pull requests, prompts, or prior work, keep separate branches, run terminal and browser validation, open pull requests, and use Agent Merge. GitHub is turning Copilot from assistant into workbench, which was always the obvious destination once agents could touch the repository. GitHub
xAI launched Grok Build in early beta for SuperGrok Heavy subscribers, with plan-review-approve workflows, clean diffs, AGENTS.md support, plugins, hooks, skills, MCP servers, parallel subagents, worktree integration, headless mode, and ACP support. Another terminal agent enters the room; do try not to trip over the subagents. xAI
Meta’s Oversight Board said Meta should move faster on deceptive AI content, especially conflict-related deepfakes, and called for stronger provenance, detection, labeling, and crisis response. Cheap synthetic media is now a trust-infrastructure problem, not a moderation footnote. Meta Oversight Board
Reuters reported that the U.S. Commerce Department cleared around ten Chinese firms to buy Nvidia H200 chips, while deliveries remain stalled as Chinese buyers face domestic pressure and security scrutiny. Export control policy, commercial demand, and national industrial strategy are all pulling on the same cable. It is, unsurprisingly, not tidy. Reuters via Yahoo Finance
NVIDIA unveiled Nemotron 3 Nano Omni, an open 30B-A3B multimodal model combining vision, audio, image, video, and text understanding for agentic workflows. The useful bit is consolidation: fewer separate perception models, lower latency, and a deployable path for document intelligence, screen understanding, and audio-video reasoning. NVIDIA
Cursor introduced configurable cloud-agent development environments with multi-repo setups, Dockerfile configuration, build secrets, layer caching, validation, rollback controls, audit logs, scoped egress, and scoped secrets. In other words: the adults have noticed that parallel coding agents need something resembling infrastructure governance. Cursor
Notion introduced a Developer Platform with an External Agents API, hosted Workers, database sync, webhook triggers, a CLI for developers and coding agents, an Agent SDK waitlist, and a unified Connections tab. Notion would quite like to be where your team data, automations, and agents meet. Subtle as ever. Notion
Thoughtworks published a useful harness-engineering piece arguing that coding agents need deterministic feedback systems: linters, tests, coverage, mutation testing, Semgrep, dependency checks, and other “sensors” that let agents self-correct. The mature pattern is not better vibes; it is tighter loops between agents and the tools that already know when code is wrong. Thoughtworks
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 Gemini-centered laptop category built around Android apps, ChromeOS, Magic Pointer, and prompt-built widgets; apparently the operating system would like to become an agent now. TechCrunch Google
Krea released Krea 2, an image model built around aesthetics, style transfer, multiple style references, adjustable influence, and controllable variation across image batches. For creative teams, the interesting part is not “prettier pictures”; it is a model designed around visual direction and iteration rather than generic prompt compliance. Krea
Anthropic introduced Claude for legal workflows with more than 20 MCP connectors across contract lifecycle management, document management, e-discovery, legal research, legal AI assistants, and public-service tools, plus practice-area plugins. Vertical Claude packages are becoming a pattern: less blank canvas, more prewired systems of record. Anthropic
Isomorphic Labs, the Alphabet and DeepMind spinout, raised $2.1 billion led by Thrive Capital to scale its AI drug-design engine and therapeutic programs. AI drug discovery is moving from “interesting platform” to capital-intensive industrial bet, which is what happens when biology meets a sufficiently large spreadsheet. Isomorphic Labs
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, with a registry planned for June. It is not a court filing, which by itself makes it unusually constructive for the AI rights conversation. The Verge
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 after closing. The frontier labs have rediscovered consulting, integration, and change management. Truly astonishing archaeology. OpenAI
Anthropic made Claude Platform on AWS generally available, bringing Claude API features, IAM authentication, CloudTrail logging, AWS billing, Managed Agents, Skills, code execution, Files API, citations, prompt caching, and batch processing into native AWS procurement and governance paths. Procurement friction is boring until it kills a deployment; then it becomes the whole story. Anthropic
Claude Code v2.1.139 added Agent View for managing background sessions and /goal, which lets Claude keep working until a defined completion condition is met. Coding assistants are becoming session managers for autonomous work, which is useful, provided someone remembers to check what “done” meant. GitHub
That will do. Try to apply this before someone converts it into a twelve-slide strategy memo.