In our own test runs, an agent building the same multi-tenant SaaS wrote about 40% less code when it started from Once UI Stack, with the same 24 checks passing.
In our own test runs, an agent building a multi-tenant SaaS from Once UI Stack wrote about 40% less code than it did starting from a bare Next.js + Once UI app. Both started from the same spec, and both passed the same checks.
The test
We wrote one spec for a small SaaS called Acme Notes. It has a landing page, sign-in, organizations with roles and invites, a blog, a signed-in layout and its own branding. A coding agent built it three times from each start:
From scratch: a new Next.js app with Once UI.
From Stack: the Stack scaffolder's output, with the AGENTS.md map it now writes.
Every run then had to pass the same 24 checks. The app has to typecheck, build with no environment variables, serve every page with its content, show no template names, ship the SQL for orgs, roles and invites, and include an .env.example. All six runs passed.
The result
From scratch
From Stack
Tokens the agent wrote (median of 3)
54k
33k
Checks passed
24 of 24
24 of 24
That is 39% fewer tokens written: the code plus the agent's own replies and tool calls. Total tokens per run came out about even, so this saves writing and review, not the cost of a run. The agent still reads the code it builds on.The from-scratch runs also built thinner versions of the same pages. They left out things Stack already has, like rate limiting, avatar upload, server-paged tables and tests.
Why the map matters
Without a map, the agent read most of the template before changing anything. Stack now writes an AGENTS.md when you scaffold. It says where the brand, SEO, copy, auth and organizations live, and it lists the exact files that hold the template's name, so the agent goes straight to them.
Caveats
These are our own test runs, not a benchmark: three runs per start, one spec, demo mode without a live database. Runs vary by around 15%, which is why we say "about 40%" and don't claim a token saving.
