This week’s AI news was stacked, so I went straight into the biggest drops that actually change how we build. OpenAI just shipped Subagents for Codex, and that’s a massive deal because it turns Codex into a real Claude Code competitor. The whole point is you can spin up multiple specialized agents in parallel, keep your main context clean, and then have Codex intelligently merge everything back into one output. For big tasks like exploring large codebases or executing multi-step feature plans, this is exactly the “team of AI workers” workflow that makes agentic coding feel fast instead of chaotic. On the model side, Minimax M2.7 looks imminent (it’s literally showing up in Design Arena), and from what I tested it’s already looking strong for front-end generation. I also covered Mistral Small 2 (Apache 2.0, faster throughput, 256k context) plus Mistral’s partnership with NVIDIA to push a frontier open-source model. Then NVIDIA GTC brought major agentic + infrastructure energy: Neotron Ultra for reasoning/agents/robotics, Nemoclaw as an open-source stack, and even DLSS 5. On top of that, Claude’s 1M token context upgrade is insane for loading entire codebases and long-running agents, and we’re also seeing a wave of “AI on your desktop” momentum with Perplexity’s always-on setup and other local agent moves.