The AGORÀ Briefing — Monday, September 28, 2026

Good morning Adam. On September 27, 2026, Anthropic opened the Claude Marketplace, a public catalog of plugins, connectors, agents and products built on Claude. Inside it you'll find the integrations from Atlassian, Google, Microsoft and Salesforce. You'll find the agentic products from CrowdStrike, Cursor, Harvey, Legora, Lovable and Snowflake. And you'll find the consulting firms Accenture, Boston Consulting Group and Deloitte. What matters most is the new way software gets published and sold. Developers build connectors with Model Context Protocol and Agent Skills. Companies selling software based on Claude request a listing in the catalog. You search for a tool, install it and update it in a few steps, just as you would in an app store. Before that date, Model Context Protocol served to link one system to the next. Now any company can use it to put its software in front of Claude's users.

Hi Adam, and good morning to everyone listening. On September 25, 2026, Microsoft unveiled the new Copilot, a super app that gathers chat, code and agents into three tabs. Microsoft compares it explicitly to Office, as The Verge tells it. The tabs are called Home, Code and Autopilot. Home is the tab that opens first, and it brings Copilot Chat and Cowork together in a single place. Inside Home sits Today, a personal screen that pulls together the important emails, the meeting invitations and the Teams conversations. Jared Spataro, who leads marketing for artificial intelligence at work inside Microsoft, says Today is the place where the user gets oriented and picks the work up again right where they left it. Code lets anyone create an app, a tracker, a dashboard or an automation, and share the result. The race now is to become the one program you work inside all day long.

Good morning to you and to everyone following along. Three researchers, Zeyan Li, Jing Peng and Jianfeng Xu, published a paper on arXiv on September 25, 2026, about the evaluation of language models. When one model grades another model's answers, the researchers measure how much the grade depends on the way the judge collects information before it decides. The judge model can work in three ways. It compares two answers side by side, or it reasons step by step, or it checks each answer against a reference answer. The ranking among the three ways shifts as the benchmark shifts, and as the judge model shifts. The trial covers four reference benchmarks and two base models with 8 billion parameters, eight combinations in all. The method the authors propose is called Backbone-Adaptive Evidence Routing. It is the most accurate across all eight combinations and it answers every case. It scores between 0.87 and 7.32 points above the best rival method. The gap between 0.87 and 7.32 tells you more than the fact that it won all eight tests.

Greetings to Adam and to everyone listening. Axel Voss sits in the European Parliament for the CDU, the same party as Ursula von der Leyen. He was among the rapporteurs on the AI Act. He told Politico.eu that the Commission moves too slowly for the speed of artificial intelligence. And he added that the rule has been applied far less than people expected. The implicit target is the European AI Office, the body charged with applying the Regulation, as key4biz reports. Voss also served as rapporteur on the files covering the metaverse and civil liability tied to AI. The law in force stands intact. Regulation 2024/1689 has been in force since August 1, 2024. The bans in Article 5 and the literacy obligation in Article 4 have applied since February 2. Criticism from politicians changes the tone of the debate, but it does not move any of these deadlines.

Good morning Adam, and greetings to everyone listening. In 1785 the British Parliament banned the export of textile machinery. The same laws also barred skilled workers from emigrating. In 1789 Samuel Slater left England disguised as a farm laborer, with zero machines and a complete memory of Arkwright's looms. In 1793 he opened, in Pawtucket, Rhode Island, the water-powered spinning mill that carries his name. London lifted the emigration ban in 1825 and the machinery ban in 1843. By that point, that knowledge had already left the country. On September 25, 2026, the South China Morning Post gave an account of a study by the think tank Carnegie China on AI talent. Export licences can stop shipping containers, patents and machine tools at the border. The skills a person carries in their head cross the border along with them. The setting in 2026 is different, but the pattern is the same.

Good morning to all of you. In March 2025 Coronis Health announced a deal with UiPath. One part of the deal was classic process automation, known as RPA. The other part was software that reads clinical documents on its own. Eighteen months later, the system has grown in a fresh direction. The company built its own software on top of the UiPath platform, and it held four things together: the automation of procedures, the reading of documents, agents that decide for themselves, and the coordination of all the work. That system now processes 100,000 reimbursement claims a week. Chief technology officer Doug Marcey calls this approach accountable intelligence, meaning intelligence that someone answers for. And he starts from a simple principle. People do the work that calls for judgment, machines do all the rest. Healthcare revenue cycle management ranks among the environments most hostile to automation. Margins are thin, the systems do not talk to each other, and the rules are dense. Buying ready-made agents is a purchasing decision. Building your own software on top of the platform is an architectural decision.