Google’s open-source Genkit now bundles conversational state, tool loops, and streaming into a single Agents API. Developers can define an agent once and run it on server or client with built‑in persistence, cutting boilerplate and accelerating multi‑turn AI app builds.
Google’s new coding‑agent skill automates a five‑stage evaluation loop, data prep, run, analyze, fix, verify, so prompt tweaks can’t silently degrade your agents. By embedding AutoRaters and independent scoring, it forces disciplined, metric‑driven improvement, catching regressions before they hit production.
The post argues that true automatic programming requires developers to assume Linus Torvalds' leadership role while AI agents act as maintainers of individual subsystems. This turns AI from a copy‑paste tool into an orchestrator of complex kernel work, accelerating progress but still demanding human oversight for overall coherence.
MoonshotAI’s Kimi K3 pushes open LLMs to 2.8 trillion parameters, activating up to 104 billion weights per token via a 896‑expert MoE. It adds a 1‑million‑token context window and built‑in vision, delivering frontier‑level coding, reasoning, and agentic performance while remaining open‑source.