OpenAI Builds Its Own Git System to Handle AI-Generated Code

Morning, {{first name | folks}}! AI is moving deeper into the infrastructure behind the internet. OpenAI has built its own Git backend to handle the surge in AI-generated code, Meta has shifted Instagram Live to an open video protocol, and Anthropic has cut response times for long AI conversations by 35%. Meanwhile, Google is expanding its AI lineup and websites are beginning to prepare for a new kind of visitor: AI agents.

OpenAI has built Pando, a new backend for Git, after the growing volume of code changes from Codex began overwhelming its existing infrastructure. It now handles nearly one petabyte of Git data every day and has cut median push latency by almost five times.

OpenAI built Pando in just 10 weeks with help from Codex and AI agents. As developers work alongside more agents and generate code at a much faster pace, even long-standing tools like Git are having to catch up.

Meta has started using Media over QUIC (MoQ), an emerging open standard, to carry Instagram Live streams on iOS and Android. After months of testing, the new protocol matched Meta’s existing system on streaming quality and reliability, allowing the switch without disrupting broadcasts for creators or viewers.

Instagram Live relied on Meta’s own RUSH protocol for nearly a decade. Moving to MoQ puts its live video infrastructure on a shared industry standard, with potential applications in lower-latency streaming and real-time interactive AI.

Anthropic has introduced Message Threads for its Messages API, giving developers a way to continue long-running AI conversations without resending the entire message history with every request. In tests involving 500,000 to 1 million tokens, the feature reduced median time to first token by around 35%.

Now in public beta, the feature is aimed at coding agents, computer-use tools and other workflows that accumulate lengthy conversations. Claude still processes the full history, so token usage remains unchanged. The improvement is in how much data developers need to send back and forth with each request.

Google introduced Gemini 4 Argon for complex tasks, followed two weeks later by Gemini 4 Carbon, its latest model focused on long-running engineering work and cybersecurity. It also unveiled Genie 4, which lets users create and steer interactive simulated worlds in real time using natural language, complete with generated audio.

Other updates include SynthID Detector, which helps identify Google's AI-generated content, and EmbeddingGemma 2, an open model for on-device multimodal search. The releases extend Google's AI work across model development, interactive experiences and content verification.

AI agents are starting to do the browsing for us, from comparing products to booking travel. That is already changing how websites operate. TIME is experimenting with versions of its articles that AI agents can read, while Etsy and Expedia are working on ways for agents to use their services.

The change raises an uncomfortable question for the companies behind the web: if an AI agent is choosing what to read, buy, or book who are websites actually trying to reach? The answer could reshape everything from online advertising to how businesses compete for customers.