How Qwida was built: every AI token, every hand-drawn line
Coding agents wrote most of Qwida's code under human review. Instead of describing that in words, this page lists every model, every session and every hour, straight from the agents' own logs, so anyone can audit it. Every image in the app was drawn by hand; the timelapses are right below.
Hecho a mano Made by hand
Minutos por pieza Minutes per piece
Trazos por pieza Strokes per piece
Arte de Yari, dibujado a mano en Procreate. Presentación de Heather, hecha a mano en Canva (unas 6 a 7 horas, estimado de ella). Unos pocos íconos de relleno (las medallas de Logros, el micrófono, el sello, el sobre y el referido) se generaron con IA en el estilo de Yari mientras ella dibuja los definitivos.
Quiénes lo hicimos Who made it
Cómo trabajamos How we worked
Context first, then orchestration. Research and decisions were written down before code, and people decided what shipped.
- A private war room repo with
HANDOFF.md, decisions and research notes, read by every agent. - Herdr: terminal panes to run several coding agents side by side.
- PioRelay: an agent to agent mailbox on the team's private network.
- Hindsight: self hosted shared memory, so a decision made by one agent is recalled by the others.
- grill-with-ui: decision interviews on a browser page, so Armando could choose fast.
- Simulated judge panels before the pitch, scored by a model that did not build the feature.
- Live link QA on every deploy, and an end to end test of the whole demo path.
Agents and their jobs
- Claude Code (Opus 5.5, Fable 5.1, Sonnet 5.5): lead, architecture, code, reviews.
- Antigravity with Gemini: research, pitch copy, integration and judge rounds.
- OpenCode with free models (Space Bunny, Muse Spark): QA, docs, feature work.
- Co-Co on GLM: red team verification.
- Marley (agent): head of the UI lane, the Familia dashboard, live QA on every deploy.
- Roman (agent): orchestration, voices and props.
Tokens by model
Tokens processed (input, output and prompt cache). Hover a bar for the breakdown. free marks free tier models.
Activity by hour (AST)
Tokens processed per hour, Thursday Oct 8 to Saturday Oct 10, for every agent on the laptop (hourly logs carry no repository; the tables below are Qwida only). Hover a column for the models used that hour.
Full ledger: repository, tool, model
Every session
One row per agent session and model. Session ids are the first 8 characters of each agent session id. Raw data: data.json.
Todo lo que usamos Everything we used
What this does and does not cover
Covered: every coding agent session on the team laptop between Oct 8 and Oct 10 in the Qwida repositories (Claude Code, Antigravity, OpenCode), read by tokscale from the agents' local logs.
Not covered: agents that ran on our two servers, and the AI calls Qwida itself makes for users (speech to text, photo reading, the week card). Those run on OpenAI and are not development work.
Cost column: what these tokens would cost at public API list prices. We worked on flat subscriptions and free tiers, so this is a measure of size, not an invoice.
Reproduce: tokscale models --json --group-by workspace,model --merge-worktrees --since 2026-10-08 --until 2026-10-10