The AI team now ships the backend too — and builds on our top model

We built the AI team because one model in a chat box is asked to be a product manager, a designer and an engineer at once, and does all three at about 70%. Splitting the work into specialists — each with its own context, its own job, its own output you can inspect — produced better apps than a single agent trying to hold everything at once.

Two things changed this week, and together they change what the team is for.

Devin: the team has a backend engineer now

The original line-up planned, directed, designed, and built a frontend. A "full-stack" result meant a UI wired to a managed backend someone else runs.

There's now a backend engineer in the room. Before the frontend engineer writes a line, Devin designs the API — the data model, the endpoints, the shape of every response — and generates a real server in Node, Python, C# or Java with a real database behind it. Not configuration against a hosted platform. A backend you own, can read, and can deploy.

The sequencing is the point. Devin's API contract is handed to Sam, so the frontend is written against endpoints that actually exist rather than endpoints the model imagined. The two halves are built against one specification, which is the difference between a frontend that talks to a backend and a frontend that talks to a backend that is there.

The coding roles moved to our top model

Team Mode was always meant to be the premium path, but its two engineers were running on the previous generation. They now build on Claude Opus 5 — Anthropic's model for complex agentic coding — for everyone on a paid plan.

The specialists who write specs and design direction stayed where they were, deliberately. Running spec-writing on the most expensive model available buys very little; running code generation on it buys a lot. The upgrade went where the difficulty is.

One subtlety we had to fix in the process: the backend engineer and the frontend engineer must be on the same model. They bind to a single API contract, and building the two halves with different models is precisely how they drift apart — one of them interprets the spec slightly differently and the seam between them breaks. Model choice is now resolved once per build and handed to both.

Sam builds feature by feature

The frontend engineer no longer emits an app in one pass either. Sam takes the PM's stories and Devin's contract, plans the features, and builds them one at a time — checking each in a live preview before starting the next. If a check fails, that feature gets fixed before anything is built on top of it.

So the team's output isn't a large generation you now have to verify. It's a sequence of features that were each verified on the way past.

What it still asks of you

The thing people notice first about Team Mode isn't the specialists — it's that you stop writing prompts.

You answer a PM's questions. You approve a direction, or ask for a different one. You look at the foundation before the rest is built on it. The work in between is not yours to specify, which is the actual difference from every builder where output quality is a function of how well you worded the request.

That remains the trade: it is slower than typing one sentence and getting an app, and it costs more, because five specialists and a per-feature check use more tokens than a single generation. Every account gets one full team run free, on the house, for exactly that reason — the claim is easier to check than to argue about.