What they did
João Bertacchi built JOBE Vox, a transcription service with Stripe billing, three environments, CI/CD and backups, using a spec-driven multi-agent process. One orchestrator agent, Claude Opus in Claude Code with autocompaction at 400k tokens, agrees goals with him, gets a plan reviewed by a plan-reviewer agent, then dispatches developer agents into isolated git worktrees. Every developer's work goes to a reviewer agent before it is merged into the branch. One PR per feature then gets CI, an automated Codex review and his own review, with fixups squashed at the end. He moved orchestration onto Paseo, which runs each lab's own CLI under agent profiles, and spread work across $20 Claude, $20 ChatGPT and $10 OpenCode Go plans. He then analysed two months of OpenCode, Claude Code, Paseo, Codex and GitHub logs.
How it went
Before the switch, orchestrator turns slowed about 8x as context grew, and after it the median stayed near 5 seconds at every size. For the same planning and building phase the orchestrator now uses about half the tokens and a third of his messages. He spent the savings on new steps: 66 percent of orchestrator tokens now go to PR, review and release work, and a typical feature uses 7 developer agents and 8 code reviewers. The Codex bot's findings per review dropped from 0.64 to 0.18.
Worth knowing
It is one developer on one project over about six weeks, and the model, the terminal tool and the orchestrator all changed at once, so he makes no productivity claim. Most rules in the orchestrator's memory came from correcting its mistakes, such as never merging without a reviewer. Cheap models handled about 207 of the code reviews.