Right. The vendors have been busy, which is rarely a sentence that improves anyone's calendar.
Cursor released a public-beta TypeScript SDK exposing the agent runtime behind Cursor desktop, CLI, and web, including local or cloud execution, streaming, cancellation, and lifecycle controls. This is the editor becoming a programmable agent platform, which is the sort of thing that sounds dull until your internal tools, CI, and review systems start wanting the same machinery. Cursor Docs
Microsoft and OpenAI amended their partnership so OpenAI can serve products across any cloud provider, while Microsoft keeps a non-exclusive IP license through 2032 and remains OpenAI's primary cloud partner. For anyone buying or building on frontier models, this changes the procurement map: OpenAI is no longer quite so Azure-shaped, which one imagines was discussed calmly in several windowless rooms. OpenAI Reuters
Google is reportedly committing up to $40 billion to Anthropic in cash and compute, with $10 billion immediate and the rest tied to milestones. Anthropic now looks less like a model lab and more like a sovereign infrastructure negotiation wearing a chatbot costume; if your roadmap depends on frontier capacity, the cloud balance sheet is now part of the product surface. TechCrunch Reuters
OpenAI released GPT-5.5 with a conspicuous emphasis on coding, browser work, long-running agents, research, and data analysis. The useful signal is not another model number for the trophy shelf; it is that the frontier fight has moved decisively toward agents that can implement, inspect, test, and keep going without needing a biscuit every six minutes. OpenAI GitHub
Mistral introduced Medium 3.5, a 128B dense model with a 256K context window aimed at instruction-following, reasoning, coding, and vision, with open weights and availability in Le Chat Work mode and Mistral Vibe remote agents. Another serious self-hostable option for teams trying to avoid renting their entire nervous system from one vendor. Mistral
Talkie-1930 is a 13B open-weight "vintage" language model trained on 260 billion tokens of pre-1931 English text, with base and instruction-tuned checkpoints on Hugging Face. It is useful for more than parlour tricks about asking a pre-war model about Hitler and asbestos stocks: the hard cutoff gives researchers a cleaner test bed for contamination, generalisation, and how much modern model behaviour is really just the web wearing a lab coat. Talkie Decrypt
AWS and OpenAI brought OpenAI models, Codex, and OpenAI-powered Bedrock Managed Agents into limited preview on Bedrock, wrapped in AWS identity, logging, encryption, PrivateLink, CloudTrail, and the usual enterprise ballast. Less glamorous than a new chatbot, which is often how one recognises useful infrastructure. OpenAI AWS
Anthropic released Claude connectors for creative workflows including Ableton, Adobe Creative Cloud, Affinity by Canva, Autodesk Fusion, Blender, Resolume, SketchUp, and Splice. That pushes Claude from "help me write the brief" toward "operate the tools that produce the asset", which is where many AI product claims finally meet something with a render queue. Anthropic
Windsurf introduced Devin for Terminal, a Rust-based local CLI agent that can share sessions with Windsurf and escalate work to Devin Cloud for VM-based testing and PR creation. The pattern is becoming clear: local agent first, cloud execution when the job needs heavier machinery or a place to make a mess. Windsurf
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 attempting to give the logs a vocabulary. Sentry
TestMu AI launched Kane CLI, a terminal-native browser automation and verification tool with support for Claude Code, Codex CLI, Cursor, and Gemini CLI, plus Playwright export and CI/CD runs. The field is learning, slowly, that generating code is not the hard part if no one can prove it works. PR Newswire
Symphony is OpenAI's open-source specification for turning project-board items into isolated Codex workspaces, monitoring CI and PR state, recovering stalled work, and moving agent-produced changes toward review. The interesting part is not the coding agent; it is the control plane around it, where teams usually misplace accountability. 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. It is another sign that vendors are racing from "chat with a model" toward "govern a process", which is considerably less shiny and rather more useful. Mistral
Chinese regulators reportedly ordered Meta to unwind its $2 billion-plus acquisition of Manus, an AI agent startup incorporated in Singapore but rooted in China. The message is not especially encrypted: agent IP, AI talent, and strategic software companies may not become portable simply because the cap table has acquired an offshore accent. Reuters
OpenAI shut down the Sora video app while keeping the API alive through September, after reported compute costs wildly outpaced revenue. Useful lesson: viral usage is not a business model when every delightful ten-second clip arrives with a small invoice from the GPU mines. OpenAI Help
DeepSeek released V4-Pro and V4-Flash preview models, with mixture-of-experts architectures, 1-million-token context windows, and availability through its API and Hugging Face. If you are evaluating open-weight infrastructure for extended coding or analysis loops, this belongs on the test bench rather than the inspirational slide deck. DeepSeek Hugging Face
Claude Code's Week 17 release put /ultrareview into public research preview, dispatching a cloud fleet of bug-hunting agents against branches or PRs and returning findings to the CLI or Desktop. High-risk merges are exactly where "several agents argued with the diff" may be a better ritual than "someone glanced at it before lunch." Anthropic
Cursor 3.2 added /multitask for splitting larger requests into async subagents, plus better worktrees and reusable multi-root workspaces. Useful for monorepos and cross-system work, assuming you have accepted that one agent was insufficient chaos. Cursor
Affirm described pausing normal engineering delivery for a week to train more than 800 engineers on agentic workflows using Claude Code, worktrees, explicit checkpoints, and automated verification. The numbers matter: 92% of engineering submitted at least one agentic PR during the week, and more than 60% of PRs are now agent-assisted. That is not a think piece; that is an operating-model change. Affirm
Google brought partner-built agents from its Agent Marketplace into Gemini Enterprise's Agent Gallery, with IT approval flows, cryptographic agent identity, Agent Gateway, and Model Armor protections. Enterprise agent sprawl is apparently no longer a risk so much as a product category, so a governed catalogue is at least the correct sort of bureaucracy. Google Cloud
xAI launched grok-voice-think-fast-1.0, a real-time voice agent model for customer support, sales, booking, and enterprise workflows, with background reasoning, tool use, structured data capture, and support for more than 25 languages. Voice agents are moving from novelty calls to operational plumbing; do try to make yours less insufferable than the current standard. xAI
Linear Agent can now connect to external systems through MCP, bringing context from tools like Granola, Glean, Notion, and PostHog into issue investigation, project planning, spec writing, and updates. Product work does not happen in one tidy database, regrettably, so cross-tool context is not a luxury. Linear
That will do. Try not to turn all of it into a roadmap by teatime.