Agent team

I build my products alone, after work. One AI coding session at a time is slow, one subscription runs out of quota in the middle of a task, and an agent that checks its own work misses its own mistakes. I also wanted to steer the work from my phone without reading every session.

An orchestrator. One Claude Code session reads the task board, writes a brief for each task, starts a worker session for it and merges what comes back. It answers the workers' engineering questions itself and asks me only about money, accounts and product direction.

Parallel workers, placed by quota. Workers are Claude Code and Codex sessions, each in its own terminal tab, spread over several subscriptions. A load balancer reads every account's remaining quota once a minute; when one runs low, it asks the worker to save its state and moves the task to an account with room.

Review by the other engine. Code that Claude writes is reviewed by Codex, and the other way round. Reviewers report problems by file and line, or say they found none; they never rewrite the author's work.

A phone channel with a gate. A relay brings my messages from my phone to the orchestrator, so I run the whole system on my computer just by chatting from my phone. Every reply passes an outbox gate that refuses internal IDs, agent names and estimates in days or weeks, so what reaches me is short and plain. An ask ledger keeps every request I make until it is answered.

Lessons it writes for itself. When something goes wrong, the orchestrator writes a dated lesson. A lesson that turns out to be a rule changes the briefs, the worker rules or the scripts.

It runs Handlez, a fantasy basketball app on the App Store, end to end: backend, app updates, store pages and marketing, videos included. It also built this site and my CV. Between 17 and 25 October 2026 it started 411 agent sessions (264 Claude Code, 147 Codex) and made about 580 commits to Handlez's backend and app.