You're here. Good. The machines have been busy, the humans have been funding them, and several regulators appear to have located their stationery.
Cognition, maker of the Devin autonomous coding agent, raised more than $1 billion at a reported $25 billion pre-money valuation, led by Lux Capital and General Catalyst. The company claims $492 million in annualized revenue run-rate and customers including Mercedes-Benz, NASA, Goldman Sachs, and Santander, which is the sort of customer list investors tend to find soothing. For anyone building with coding agents, the lesson is plain: the market still believes there is room for independent agent companies next to the frontier labs, at least until the frontier labs notice properly. TechCrunch
xAI launched Grok Build, an early-beta terminal coding agent for SuperGrok and X Premium Plus subscribers, with plan-and-approve workflows, repository conventions, skills, hooks, MCP, parallel subagents, worktree integration, headless scripting, ACP, and a grok-build-0.1 API model. In other words, another lab has discovered that serious coding agents require a small operating system of process around the model. Sensible, if somewhat belated. xAI Docs
The Vatican published Pope Leo XIV's first encyclical, Magnifica Humanitas, focused on safeguarding the human person in the age of artificial intelligence. It covers dignity, labor, accountability, data ownership, environmental costs, automated decisions, and concentration of technological power, because apparently AI governance has now moved from product councils to papal doctrine. Companies treating AI ethics as a slide in the sales deck should note that the conversation has escaped the building. Vatican Simon Willison
Alibaba's Qwen team released Qwen3.7-Max, a closed-weight, API-only reasoning and agentic model with a 1M-token context window and Anthropic-compatible API support. Reports describe long autonomous runs, including roughly 35 hours of kernel optimization and more than a thousand tool calls, which is either impressive engineering or a request for adult supervision. For agent builders, the important bit is endurance: sustained execution is becoming a model capability, not merely an orchestration trick. VentureBeat MarkTechPost
Claude Code v2.1.152 added /code-review --fix, applying review findings directly to the working tree, and brought /simplify back as an alias for that flow. It also added tool-deny frontmatter, /reload-skills, richer hooks, plugin marketplace controls, and fallback-model behavior, because apparently one coding agent now needs a small compliance department. GitHub Anthropic Docs
OpenAI and Thrive described Tax AI, a Codex-driven improvement loop deployed with Crete's network of 30-plus accounting firms. The useful pattern is not tax preparation, unless you are particularly unlucky; it is turning practitioner feedback and production traces into eval tasks so an agent can improve against the actual work. OpenAI
Slack now lets paid-plan users add AI-generated answers and channel summaries inside Workflow Builder, drawing on prompts, workflow variables, files, canvases, or channel history. This moves Slack AI from "find things for me" into no-code automation, which is where team tools become interesting and occasionally dangerous. Slack UC Today
Simon Willison argues OpenAI and Anthropic have found product-market fit less through consumer chatbots than through expensive enterprise coding and general-purpose agent usage. It is a useful synthesis of the week's anxiety about token costs, lab economics, and whether agent usage has finally become a business rather than a demo with invoices. Simon Willison
OpenRouter raised a $113 million Series B led by Alphabet's CapitalG, with NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, Andreessen Horowitz, and Menlo Ventures participating. Its reported 25 trillion tokens per week and 8 million users suggest enterprises are warming to model routing rather than swearing eternal loyalty to one provider. Practical. Also deeply inconvenient for vendor lock-in decks. TechCrunch Business Wire
Pulumi published a practical guide arguing that the model is often less important than the repo-specific harness around it: layered instructions, hooks, path-scoped skills, language-server-backed navigation, MCP tools, and exploration subagents. Translation: stop polishing prompts as if they are heirloom silver and start maintaining the system that lets agents work reliably. Pulumi
President Trump canceled a planned executive-order signing that would have created a voluntary government framework for pre-release testing of frontier AI models. Reporting cited industry resistance, arguments over testing timelines, and concern that the order could slow U.S. competitiveness against China. The result is the usual: unresolved oversight, louder policy fights, and everyone insisting they are the responsible party. Ars Technica
Illinois advanced SB315 in a 52-5 Senate vote, targeting large frontier model developers with transparency frameworks, third-party auditors, and catastrophic-risk reporting. It is modeled on California and New York efforts, because state-level AI governance is apparently now a franchise system. WTTW
Microsoft Developer published an Agentic-Agile methodology: specs in the backlog first, contract-driven execution, incremental delivery, persistent repo instructions, governance as acceptance criteria, and measurement of agent collaboration quality. Sensible teams are moving from "prompt harder" to actual engineering process. Extraordinary development, really. Microsoft Developer GitHub
LaunchDarkly introduced AgentControl, an operational layer for changing prompts, models, parameters, and tools at runtime, with evals, tracing, progressive rollouts, and adaptive triggers. Feature flags have arrived for the things that make decisions on your behalf. This is the sort of unglamorous control plane that determines whether agent deployments survive contact with production. LaunchDarkly LaunchDarkly Docs
Runway launched Aleph 2.0, a video editing model inside its new Edit Studio product, with localized edits, image-level control, multi-shot editing, and up to 30 seconds of 1080p video on paid plans. Video generation is steadily becoming video manipulation, which is the useful part and, inconveniently, the dangerous part. Runway Runway
SageMaker AI real-time inference endpoints now expose an OpenAI-compatible API path, so developers can use the OpenAI SDK, LangChain, or Strands Agents by changing endpoint URLs instead of rewriting clients. Compatibility is not glamorous. It is, however, how migrations actually happen. AWS AWS Blog
Google introduced Agent Executor, an open-source distributed runtime standard for agent execution, resumption, and deployment, alongside Agent Substrate for high-density agent infrastructure on Kubernetes. The point is durable execution, session consistency, connection recovery, secure sandboxing, and mixed deployment models. Agents are being treated less like chat sessions and more like distributed workloads, which was inevitable once they started running long enough to require babysitting. Google Cloud Google Cloud
Cohere released Command A+, a 218-billion-parameter sparse mixture-of-experts model with 25 billion active parameters, 128K input context, multimodal inputs, tool use, reasoning outputs, and support for 48 languages. The Apache 2.0 license and quantized weights designed for one B200 or two H100s make it unusually relevant for companies that want capable enterprise AI without sending every internal document elsewhere. Sovereignty is much more persuasive when it fits in the rack. Cohere Hugging Face
Figma introduced a native AI design agent in beta that works directly inside Figma Design. It can generate and edit layouts, automate bulk design tasks, use components, variables, and team context, and help with comment summarization and iteration. The important part is not that AI makes pictures; it is that it works where the design system already lives. Figma TechCrunch
Cursor 3.5 moved Automations into the Agents Window and added support for recurring work across multiple attached repositories, or with no repository at all. That turns agent work from a one-off coding session into a scheduled operations surface across code, Slack, analytics, finance, and customer-health workflows. Yes, someone will automate the wrong thing. No, that does not make the feature unimportant. Cursor
That's the lot. Do brief someone before they confuse infrastructure with strategy. Again.