Good morning, and welcome back to Techzip AI Newsletter.

TL;DR: DeepSeek sketches a system for coordinating parallel AI agents; GPT-6 Astra edits Qwen3.5-0.8B’s weights to break a reasoning loop; and OpenAI’s Pando scales Git for coding agents. Elsewhere, OpenAI tests flexible data-center power and a combined ChatGPT Chat/Work experience, while Anthropic warns future models may approach AI self-improvement and adds stateful threads to the Claude API. The issue also covers AI evaluation, agent memory and search, AI app economics, Grok Bot’s serving costs, Apple’s potential AI-app revenue and Sam Altman’s argument for cloud-based DRAM.