The AGORÀ Briefing — Saturday, August 22, 2026

Good morning, everyone. On August 20, 2026, eSafety, the Australian online safety regulator, declared that Roblox keeps exposing minors to contact with unknown adults. The regulator acts under the Online Safety Act, the law in force that governs digital platforms. Here the investigation becomes public enforcement. Tests run in 2026 show measures too weak to block contact between adults and children under 16, as The Verge documents. Roblox stays among the most widespread platforms for minors, and that turns the case into a precedent for the whole sector. The rule applies to anyone who hosts those children. Whoever writes the obligations decides who answers.

Thanks, Adam. On August 18, 2026, Cursor announced Origin, a code-hosting platform built to contest GitHub's turf. The launch lands while the incumbent goes through a wave of documented outages. Cursor, now officially part of SpaceX, aims at the heart of the developer workflow. The message is direct: whoever controls the repository controls the relationship with the engineering team. The repository sees every line of code, and from there other services spring, from review to automation all the way to AI agents. Origin replicates GitHub's core functions, from browsing to editing all the way to pull requests. Cursor also promises native agent functions and a wider app ecosystem. The code layer opens back up to competition.

Hi, Adam. ComponentBench, published on August 18, 2026, and accepted at COLM 2026, measures component-level failures in computer-use agents. The suite covers 97 canonical UI components, instantiated across 2.910 tasks verified programmatically. It fills the middle layer that current benchmarks leave uncovered, between long-horizon workflows and atomic GUI-grounding tests. Each task arrives with clean human reference trajectories. That allows success and efficiency to be measured in the same run. The result is a dual metric. An agent completes the task with a step count far above the human one. Separating success and efficiency exposes the agent that looks robust and is simply slow.

Thanks, Adam. In 1968 the Federal Reserve adopted the MPS model, built with MIT and the University of Pennsylvania. The mechanism replaced human judgment with the equation, inside a rising inflation. That model promised stable forecasts and delivered systematic errors during the stagflation of the Seventies. It held dozens of behavioral equations, and its elegance hid a flaw: inflation expectations ran faster than its estimates. Monetary institutions love tools that reduce individual responsibility. An equation absorbs blame better than a governor. Today the AI central bank enters the reaction function with the same pattern. A tool calibrated on one regime collapses when the regime shifts.

Hi, everyone. The FCC killed the gigabit goal, and the consensus reads the move as a surrender. The consensus watches the wrong figure. Chairman Brendan Carr abandoned the Biden-era targets, which aimed at gigabit downloads and half-gigabit uploads, as The Verge reports. Analysts fix on the regulatory target, an indicator that lags the market. The figure that matters is the cost per megabit delivered. The federal standard rose from 25/3 to 100/20 in 2024, and the market was already racing past that threshold. A cliff event will make gigabit speed the market standard by 2032, regardless of any federal target. The cost trajectory describes the future.

Thanks, Adam. Economic D-Day marks the point where an economic trajectory diverges in a measurable way from the formal commitments made. A new study applies it to the European climate. The European Union promised to cut greenhouse gases by 55 percent versus 1990 levels by 2030. The study checks the compatibility between current trends and this promise. The authors apply deep learning models to socioeconomic and sectoral data from EU27 through 2023. The evidence shows a divergence that grows predictably with sectoral inertia. The gap becomes measurable, beyond rhetorical. A measure with a declared sample and method shrinks the space for discretion. Here the data comes before the interpretation.