You do not need a different AI tool for every job. One team of seven Codex agents covers planning, building, testing and review for the work I do. The same names and the same setup work in every repository.
One team is enough
Each task gets its own worktree, so agents never share a half-finished branch. The team is the same in every repository. Only the task changes.
The seven agents
| Agent | Owns | Model and reasoning |
|---|---|---|
| codex-lead | Architecture, planning, assigning work, coordination | GPT-6 Astra, extra-high reasoning |
| codex-feature | Full-stack features, APIs, databases, billing | GPT-6.1 Sol, high reasoning |
| codex-ui | Front-end code, component library, responsive layout, accessibility | GPT-6.1 Sol, medium reasoning |
| codex-infra | Infrastructure as code, containers, CI/CD pipelines | GPT-6.1 Sol, high reasoning |
| codex-integrations | Source hosts, payments, identity, external APIs | GPT-6.1 Sol, high reasoning |
| codex-qa | Automated, browser, integration, failure and performance tests | GPT-6.1 Sol, high reasoning |
| codex-review | Independent review: security, concurrency, billing, architecture | GPT-6 Astra, extra-high reasoning |
The lead plans the work and hands each piece to one builder. The four builders write the code. QA proves it works. Review checks it without having written any of it.
Ship a usage export for customers
The rules
- The lead assigns and coordinates. It does not write the feature code it hands out.
- Builders own their branches and pull requests. When two changes touch the same files, the builders talk before either merges.
- Review stays independent. Review never reviews code it wrote, and it does not edit the code it reviews.
- Billing and security need real tests. A passing run is not enough. The tests must fail when the rule breaks.
- Trivial work goes to a small model. A rename or a typo fix does not need an eighth agent. The small model is the fallback.
What it costs
The team runs two GPT-6 Astra agents and five GPT-6.1 Sol agents. GPT-6 Astra is used only where judgement matters most: the lead plans the work and review checks it. The four builders and QA run on GPT-6.1 Sol, four at high reasoning and the UI agent at medium.
Three habits keep the bill down:
- Match the model to the task. Most of the work does not need GPT-6 Astra.
- Set reasoning effort per task. Medium is enough for UI polish. High is for logic, billing and infrastructure.
- Share one team across repositories. Do not build a separate team for each project.
When to add another harness
Codex alone is enough for this work. Add another agent tool only when a task needs something Codex cannot do. Until then, one harness with one team is simpler to run, review and pay for.
The coordination skill this team runs on is published as zs-orchestrate.
Try it
Start with the lead on a small, real task. Let it split the work, review what comes back, and measure the cost after a week. Then decide whether the team needs more agents.