Patreon tweaks algorithms as AI agents reshape web infrastructure

Patreon announced changes to its discovery algorithm this week, shifting away from favoring large creators toward comparing post content directly—a move CEO Jack Conte framed as building "a better network" in contrast to platforms that have "enshittified."

The update will analyze post topics, style, and themes rather than recommending creators based on follower networks, which historically favored established accounts. Patreon is also rolling out Clips, a tool that automatically extracts short-form video moments from longer content for sharing on and off platform.

The changes arrive as infrastructure for AI agents continues to mature. Argentic launched an L402 Lightning toll booth that lets AI scraping agents pay per request in Bitcoin—10 satoshis (~$0.001) for access, settled in under 40 milliseconds. Agents receive a one-hour session token after payment, no accounts or API keys required.

Meanwhile, OzBrain debuted as a shared knowledge layer that multiple AI agents can read from and write to. The system structures information so agents "read only what they need," according to the company, and integrates with Claude, ChatGPT, and Cursor through a custom MCP connector.

On the hardware side, Micron announced a $10 billion research hub in Boise focused on memory and AI. The facility adds to a broader infrastructure buildout as companies race to support increasingly compute-intensive workloads.

Controversy continues around AI training data acquisition. A 404 Media investigation revealed Amazon is buying rare books, scanning them for training data, and destroying the physical copies at a Las Vegas warehouse. Employees at the facility, designated VGT3, cut book bindings to speed scanning. The operation targets rare titles—often foreign-language works with limited circulation—that contain "content as a bunch of words strung together," as one bookseller described it.

The practice raises questions about fair use at scale and permanent loss of physical artifacts, though Amazon has not commented on the operation.

In the agent tooling space, several projects launched this week aimed at making autonomous systems more practical. OneCLI, a YC-backed platform, provides sandboxed agent environments for teams with centralized policy enforcement and credential management. Seed takes a minimalist approach—a single Python file that connects a language model to one tool (bash exec) and a self-modifying prompt file.

Developers are also standardizing around AGENTS.md files, similar to README conventions, to help coding agents understand codebases. The format is gaining traction across Codex, Amp, and Cursor, with a feature request pending for Anthropic's Claude Code to adopt it.

The proliferation of parallel agent workflows is pushing architectural changes. As one developer noted on their blog, a team running 20-100 agents in parallel can generate 500 commits and 200 pushes daily—volumes that previously required hundreds of engineers and drove Uber's microservices architecture. "The modularity of your codebase determines how many agents you can run in parallel," they wrote.

Elsewhere, Meta shipped a Mac app for its AI assistant with window-sharing capabilities and dictation across all apps. The assistant now integrates directly with Instagram, Facebook, and Google Workspace for business users.

Emacs 31.1 is scheduled to release on August 24, and Genki launched the Manta controller on Kickstarter—a gamepad with a 2.9-inch screen, tri-stage triggers, and analog face buttons priced at $189 for early backers.