The AGORÀ Briefing — Thursday, August 27, 2026

Good morning, Adam. On August 2, 2026, the EU AI Act provisions on high-risk systems become applicable across the European Union. The phase shifts from adoption to enforcement. The regulation took effect on August 1, 2024, with a staggered rollout. The prohibited practices apply starting February 2025. The obligations on general-purpose models follow starting August 2025. Now it's the turn of Annex III, the systems classified as high-risk. This category covers uses in employment, credit, biometrics, and access to essential services. The classification sets the intensity of the obligations. A system that falls inside it activates the entire documentation perimeter. The schedule has been known for months. Today it becomes a fixed point.

Thanks, Adam. Two independent sources published in 2026 converge on a structural signal. Entry-level jobs exposed to AI are the part of the market where the transformation shows up first. On Morning Edition, NPR reports that new graduates struggle to land their first hire, and they attribute the difficulty to artificial intelligence. The Federal Reserve Bank of New York places the unemployment rate for recent graduates, people between 22 and 27 with a bachelor's degree or higher, at 5.7 percent in June. The overall workforce sits at 4.1 percent. Irene Chang, 21, finished industrial and systems engineering at Georgia Tech in May. She estimates roughly 450 applications and 19 interviews, with zero offers. These are two distinct populations. The first door weighs more than all the rest.

Hi, Adam. On April 23, 2026, Anthropic made memory for Claude Managed Agents available in public beta. The agents learn from every session through an optimized memory layer. Anthropic describes it as tuned on internal benchmarks for long-running agents that improve from one session to the next. The agents share the information they gather with each other. The detail sits in the architecture. The memory mounts directly onto a filesystem, so Claude uses the same bash and code execution capabilities that make it effective in agentic tasks. With file-based memory, the models save richer and better-organized memories. The system becomes more selective about what to keep. The move from beta to production-ready stays open.

Thanks, Adam. On August 25, 2026, NVIDIA announced Jetson Orin Nano 2, a robotics computer designed to redefine entry-level edge AI. NVIDIA moves the power of frontier models directly toward physical devices. We're talking about generative inference running on the robot, on the camera, on the industrial arm. According to the official release, the new module delivers double the inference performance of its predecessor, in the same form factor, and consumes 40 percent less energy at equal performance. Over 3 million developers already build on NVIDIA's robotics stack. Cognex, Doosan Bobcat, and Matic rank among the first to adopt it. The same form factor matters: those who have already designed the hardware can upgrade without redesigning the product.

Good morning, Adam. The precedent is 1914. A few hours after the declaration of war, the British ship Telconia moved up the North Sea and cut the five telegraph cables that linked Germany to the world. The mechanism was simple: control the flow. Berlin lost direct access to its own embassies and its own markets, while London read the residual traffic. Whoever controlled the transmission controlled the interpretation. A century later, the structure stays identical. Whoever controls the cables controls the information. And today the information includes the entire economy of artificial intelligence, from trained models to inferences distributed across continents. Nearly 100 percent of intercontinental internet traffic passes through undersea cables. The stakes have changed. The Achilles' heel stays the same.

Thanks, Adam. In 2026, Salesforce published the Agentic Enterprise Index, a report that aggregates real usage data from the Agentforce platform. The average number of agents activated per organization grew by nearly three times over the fiscal year. The average time from creating an agent to using it fell by 53 percent. Let's read it for what it is: a market benchmark, an aggregated snapshot of enterprise AI in production. Calling it a case study would be improper, because the names of the individual companies are missing. It stands as a measure of where the bar of the possible has moved. The distinction between announced data and verified data stays the heart of the reading. The date of the report matters, and we state it openly. This 2026 snapshot stays current for anyone designing today.