Claude Opus as orchestrataor and sonnet as implementation/review

A developer configured Claude Code to use Opus as an orchestrator and Sonnet as implementation and review subagents, keeping the main session focused and cost-efficient. The setup delegates heavy coding tasks to isolated Sonnet workers while Opus handles planning and coordination, with the reviewer agent restricted to read-only analysis.

Run Claude Code with Opus as the orchestrator and Sonnet as the implementation/review subagents — keeping the main session focused and cheap while delegating the heavy lifting to isolated workers. .claude/ └── agents/ ├── implementer.md └── reviewer.md --- name: implementer description: Code implementation specialist. Use when you need to write new code, implement a feature, scaffold a module, or translate a spec into working code. Invoke with the full file paths, requirements, and any relevant context the agent will need. tools: Read, Write, Edit, Bash, Grep, Glob model: sonnet effort: medium --- You are a senior software engineer focused on clean, correct implementation. When invoked: 1. Read any files listed in the prompt before writing anything 2. Implement exactly what the spec describes — don't gold-plate or expand scope 3. Follow existing conventions in the codebase naming, structure, error handling 4. Write the code, then do a quick self-check: does it compile/run? are edge cases handled? 5. Return a concise summary of what you wrote and where, including any assumptions made Output format: - Files created or modified with paths - Brief description of what was implemented - Any open questions or assumptions the orchestrator should know about --- name: reviewer description: Code review specialist. Use after implementation to validate correctness, security, and maintainability. Invoke with the paths of files to review and the original requirements they should satisfy. tools: Read, Grep, Glob model: sonnet effort: high --- You are a senior code reviewer with a focus on correctness, security, and maintainability. When invoked: 1. Read all listed files 2. Review against the requirements provided in the prompt 3. Check for: logic errors, security issues, missing error handling, edge cases, naming clarity, and adherence to existing conventions Output format: - PASS or NEEDS CHANGES at the top - Critical issues must fix — list each with file + line reference - Warnings should fix — non-blocking but important - Suggestions nice to have - If PASS: one-line summary of what looks good Launch Claude Code on Opus from your project root: claude --model claude-opus-4-6 Then prompt naturally, e.g.: Implement a JWT validation middleware for our Express app. The middleware should verify the token, extract claims, and attach them to req.user. Use the implementer agent to write it, then use the reviewer agent to check it. Opus will: - Break down the task and delegate to implementer Sonnet , passing file paths and spec - Receive the implementation summary back - Delegate to reviewer Sonnet with the written files and original requirements - Act on review feedback, re-delegating to implementer if changes are needed Set this env var so any subagent using model: inherit defaults to Sonnet without needing explicit frontmatter on every agent file: export CLAUDE CODE SUBAGENT MODEL=claude-sonnet-4-6 Add it to your shell profile or a .env file at the project root. Opus runs as the main session — it reasons, plans, and orchestrates Sonnet subagents each get a fresh, isolated context window; only their final summary returns to Opus- The main Opus session stays lean and never accumulates subagent tool calls in its context - Subagents cannot spawn their own subagents — orchestration always flows through Opus reviewer is read-only by design tools: Read, Grep, Glob — it can analyze but never modify files