← sources

My Claude Code subagent setup: orchestrator, doers, reviewer, blind tester. What would you change?

source: reddit (reddit:ClaudeCode)published 2026-10-02 02:12ingested 2026-10-03 02:40

retained text

I've been running Claude Code as an orchestrator over subagents on a mid-size web app for a few weeks. The board in the screenshot is a local page that tracks progress and everything waiting on me.

**Roles**

\- **Orchestrator (Fable)**: briefs the agents, reads their summaries, runs the full test suites. It doesn't write code itself.

\- **Doer (Opus)**: makes one scoped change and runs its own tests.

\- **Reviewer (Opus)**: read only, reports findings.

\- **Blind tester (Sonnet)**: told only what the feature should do, never how it was built. Runs one yes/no check in a real browser and ends with "ready" or "not ready".

Nothing ships until the tester says ready. Agents never touch git; I make every commit.

**Lessons learned**

\- Max 3 agents at once. With 5 or more we hit the usage limit and they all died mid task.

\- Never point several agents at the same files.

\- Only one agent drives the browser, or they close each other's tabs.

\- The cheap model tests; the expensive model only fixes what failed.

**Questions**

1. Is a blind tester worth it, or would you fold it into the reviewer?
2. How do you pace parallel agents over a long day?
3. What would you cut or change?

PS: These notes are specifically Max(100$) package.

https://preview.redd.it/f38ym8l4qysh1.png?width=2125&format=png&auto=webp&s=271e395231e0c28ee90fec6f46c6144af12a493d

document fetched 2026-10-03 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1wvhzr2/my_claude_code_subagent_setup_orchestrator_doers