The AGORÀ Briefing — Thursday, September 17, 2026

Good morning, Adam. On September 16, 2026, the Court of Appeals for the Ninth Circuit filed its opinion in Doe versus GitHub, case number 24-7700, and it upholds the partial dismissal decided at the trial level. The file comes from the Northern District of California, from Judge Jon S. Tigar, on interlocutory appeal. The case was argued on February 11, 2026, in San Francisco, and the opinion is written by Judge Eric D. Miller. In the dock are GitHub Copilot and Codex, tools that write code using language models trained on millions of software projects hosted on the platform. The decision separates two levels that boards of directors treat as one: who is allowed to sue, and which rule decides the case in favor of whoever sues. On the first level the programmers win, with standing recognized under Article Three. Risk, starting today, gets measured on two separate lines.

Hi Adam. Four researchers filed a study on arXiv on September 16, 2026: eight pages and three figures, titled TuiML, Machine Learning for AI Agents. The names are Nilesh Verma, Nick Lim, Albert Bifet and Bernhard Pfahringer. The thesis is sharp. Libraries like Weka and scikit-learn were born for human programmers, while language agents recall them from memory and write code. From this difference in audience the abstract draws three effects: the current approach hides what a library offers, it pushes errors off to runtime, and it loses the experimental state from one turn to the next. Reproducibility in machine learning has been debated for years in replication studies. Here the question arrives upstream, inside the interface between agent and tool. The filing is recent and the public version remains v1, so independent review still has to arrive.

Hello. On September 16, 2026, Google opened early access to the Google Home MCP server, as TechCrunch reports. Access stays reserved for subscribers to the Google Home Premium Advanced plan, 20 dollars a month in the United States. The server talks with any agent compatible with the Model Context Protocol: Claude, ChatGPT, Hermes, OpenClaw and Google Antigravity. Coverage includes Nest devices, doorbells, thermostats and hardware labeled Works with Google Home or Matter. The user grants permissions with a login, inside a Google Cloud project. From that moment the agent reads the event history, looks at the camera summaries and commands the lights in the house. The announcement talks about convenience at home. The commercial substance sits elsewhere: a standard bridge between Google's installed base and third-party agents, rivals included. The protocol goes on the price list.

Good morning to everyone listening. Today's number comes from the Google Cloud blog, which on September 17, 2026, published the story of SeaVerse, a gaming startup born out of SeaArt. The company claims a cut in infrastructure costs of up to 60 percent after moving to Google Kubernetes Engine with GKE Agent Sandbox. The post carries two names at the bottom: Zongyun Hu, COO of SeaVerse, and Tinsley Shi, Product Manager at Google Cloud. Customer and vendor produced the document together, and it is the single public source for that figure. Anyone reading that percentage is reading a statement from the seller, on the seller's own channel. Underneath there is a real platform. People who use it describe an idea in words and get a light game or a chat with a character. Each creation becomes a separate workload, and it has to start instantly. It remains a hypothesis to verify, just like any commercial claim.

Good to be with you. Fewer than one organizational code of conduct in ten deals explicitly with artificial intelligence or the ethics of technology. The figure comes from the report the advisory firm LRN published on September 16, 2026, based on 2,000 people employed full time. Almost all of them say their employer asks them to certify or accept that document. Coverage is total, and the content stays silent precisely about the tool that has entered daily work. Almost one person in five adds that their own code lacks practical guidance. The people who updated those texts covered the legal obligations and left the operational part outside, while HR platforms fold AI features in at enterprise scale. On the same day a federal court redraws the risk of the tools that write code, the document every hired person signs stays mute.