Claude Code subagents need real contrast

Claude Code's subagents are a coordination primitive, not a management miracle. If you want disagreement, you have to engineer it instead of assuming delegation creates it.

• –Anthropic describes subagents as specialized assistants with separate context windows, custom prompts, and tool access, so the isolation is real but only at the workflow layer.
• –If the orchestrator gives every subagent the same brief, the outputs will converge. Distinct perspectives need distinct goals, rubrics, and permissions.
• –A better pattern is planner, skeptic, builder, and reviewer, then a synthesis pass that compares contradictions instead of smoothing them out.
• –Product Hunt sentiment already hints at the gap: Claude Code is praised for strong context handling and clean output, but some users still want more autonomy than the product naturally delivers.