The Studio
A software studio run by twelve Claude Code agents, where nothing reaches a client until I approve it.
- Client
- Own product
- Year
- 2026
- Role
- Design, architecture, engineering
- Stack
- TypeScript, Bun, SQLite, Claude Code, Discord API
The problem
Freelance work is a funnel of small decisions: which job deserves a bid, at what price, with what proof. Done by hand, most of the day goes to triage and writing instead of building.
What I built
Twelve Claude Code agents, each with one job and its own playbook: scout, researcher, bid strategist, controller, skeptic, solutions architect, designer, proposal writer, engineer, QA, DevOps and client success. They debate every bid in a council, draft proposals and live prototypes, and hand won work to a parallel coding crew. Every outward action waits for my approval in Discord.
The outcome
In five days the studio shipped 14 repositories with about 245,000 lines of code, about 3,800 automated tests and CI on every repository. Among them: PayLab ID, a payments workbench with 580 tests, and DocExtract Lab, invoice extraction at a field-level F1 of 94.
How it works
- 01Job ina post arrives through Discord or a browser button
- 02Scoutscores fit, client quality, competition and risk
- 03Councilstrategist, controller, skeptic and architect debate the bid
- 04MeI approve every proposal, reply and deploy
- 05Crewa parallel coding crew builds from a spec
- 06QAtests against acceptance criteria until it ships
Details
- Per-role model routing: the strongest model where judgement matters, the fastest for triage.
- A PreToolUse safety hook keeps agents away from the logged-in browser.
- Budget-aware throttling shrinks councils before the usage limit is reached.
- Automatic Claude-to-Codex failover keeps the crew building.
- A weekly retrospective rewrites the 13 versioned playbooks from what won and what lost.
Results
- 14repositories in 5 days
- ~245klines of code
- ~3,800automated tests
