The AI Architect Briefing
Coding Models and Open Weights
The month opened on two fronts: a scramble for the AI coding market, and a notable open-weight release that keeps capable models within everyone’s reach.
Models
Microsoft and Google moved to challenge Anthropic and OpenAI in AI coding models, the category Claude Code and Codex had come to define. Google’s Gemini 3.5 Flash, generally available since I/O, is now the default across the Gemini app and AI Mode in Search, pushing fast, cheap inference to the front of the product.
Tools and frameworks
At BUILD, Microsoft detailed the direction for its Agent Framework: an agent harness, hosted agents, and a CodeAct execution model, consolidating AutoGen and Semantic Kernel into a single supported platform. The theme is fewer competing abstractions and more production-ready foundations.
Standards and open source
Google DeepMind released Gemma 4 12B, an open-weight, encoder-free multimodal model under Apache 2.0. Capable open weights staying current matters: it keeps a credible alternative to closed APIs alive for teams that need control or lower cost.
Money and infrastructure
The coding-assistant race is as much about lock-in as capability. Whoever owns the developer’s coding agent gains a durable channel into the rest of the stack, which is why every major lab is now fighting for it.
What I am watching
AI coding is the beachhead. The tools developers adopt this year shape which ecosystems they build on next, so watch coding-agent adoption as a leading indicator of platform gravity.