AI is moving from generation to review

Shopify CTO Mikhail Parakhin said the company now spends more effort on critique loops, PR review, testing and deployment stability than on generation itself. That is the right way to think about it.

The point is simple. AI can write more code than before. The problem is that more code also means more bugs, more failed tests and more pressure on delivery pipelines. Parakhin's team is pushing expensive models into review, while keeping generation relatively cheap.

That maps neatly to construction delivery. If you automate too early and skip the control points, you just create a faster mess.

OpenAI's GPT-5.5 launch bundled a much bigger Codex update. Browser control, Docs, Sheets, Slides and auto-review all moved into the same workspace direction.

That matters because it shows where these tools are heading. They are no longer just code assistants. They are becoming work agents that can move across documents, spreadsheets and web apps.

For construction teams, that means the useful unit is not the model. It's the stack around it. The review loop. The permissions. The handoff into the tools your team already uses.

Qwen 3.6 27B landed as a serious open coding model, with strong benchmark claims and broad local support across vLLM, Ollama and llama.cpp. Community reaction was immediate because the model looks good enough to compete on real tasks, not just on slides.

But, the bigger story is what that does to cost and deployment choices. If an open model can get close enough for a lot of office and back-office work, the default assumption shifts. You don't have to send everything to a frontier API.

That has obvious implications for construction firms handling sensitive project data.

Google used Cloud Next to push TPU v8, Gemini Enterprise Agent Platform and Workspace Intelligence. OpenAI answered with image generation, Codex upgrades and a stronger work-agent story.

This is all part of the same shift. The winning products are no longer just better chats. They are controlled systems for doing work. That is much closer to how construction teams actually operate.