The AI Architect Briefing
Agents Become Products
This week the agent story moved from “how to build one” to “here is the product,” even as the tooling underneath split into two design philosophies.
Models
OpenAI moved toward ChatGPT Work, an agent-based product powered by Codex and GPT-5.6, packaging agents as a shipped product rather than a raw API. GPT-5.6 itself stayed in a staged, limited rollout, a sign of how cautiously the most capable models are being released.
Tools and frameworks
Anthropic’s Claude Agent SDK, with native Model Context Protocol support and subagents in TypeScript and Python, made multi-agent systems more approachable. At the other end of the spectrum, Hugging Face’s smolagents pushed a code-first, minimal-abstraction runtime of roughly a thousand lines, where the model writes and runs Python in a managed sandbox.
Standards and open source
Code-first runtimes are gaining ground against heavy orchestration layers. The appeal is transparency and control: less framework magic, more plain code you can read and debug.
Money and infrastructure
Packaging agents as products, not APIs, changes go-to-market. It moves the value from raw tokens toward workflows, integrations, and the trust needed to let software act on a user’s behalf.
What I am watching
The real design debate of the year is heavyweight orchestration versus lean code-first runtimes. Watch which approach teams reach for when they move from demo to production.