Subagent-Driven Development

Execute a plan by dispatching fresh subagents per task, with the lead doing inline review (spec compliance + code quality) after each task. Independent tasks can be dispatched in parallel; tightly coupled tasks run sequentially. For fixes, dispatch a fresh implementer with the fix prompt and prior-task context.

Core principle: Fresh subagent per task + inline review by lead + parallel fan-out when tasks are independent = high quality without wasted ceremony.

Dispatch mechanism:

• Claude Code: Agent tool (one call per subagent; multiple calls in one turn run in parallel).
• opencode: Task tool (one call per subagent; multiple calls in one turn run in parallel). The built-in general subagent is suitable for most implementer work; @mention also works for manual invocation.
• Codex: spawn_agent(agent_type="worker", message=<filled prompt>) (one call per subagent; multiple calls in one turn run in parallel). Keep the returned agent ID, send_input(target=<agent-id>, message=...) feeds follow-ups (the closest thing to Claude Code's resume), wait_agent(targets=[<agent-id>]) blocks until the agent finishes, and close_agent(target=<agent-id>) frees the slot. Requires multi_agent = true in ~/.codex/config.toml (see skills/using-razorback/references/codex-tools.md).
• Gemini CLI: invoke_agent(agent_name="generalist", prompt=<filled prompt>) (parallel by default; set wait_for_previous: true only when you need a call serialized behind earlier ones). Resume is not available — route fix rounds via a fresh invoke_agent call with the fix prompt and prior-task context. Subagents cannot recursively dispatch other subagents, so all worker dispatch happens from the lead session.

If the harness supports per-agent model or reasoning selection, apply the plan's Model Routing tier when dispatching. If it does not, use inherit and note that in the task report.

When to Use