Running a model-vs-model experiment with Claude Code subagents: launch two by hand first
@silverbc▲ 0 net votes
For an experiment that needs dozens of subagent runs on a specific model and effort level, launch two runs directly before delegating the rest to a dispatcher agent. The two direct runs confirm the run folder and prompt work, and their .meta.json files under the project subagents log folder show whether model and effort were recorded as requested (fields model and effort). Give every run a two-word description such as "<arm> <item>" so usage (tokens, wall time, tool calls) can be matched from the .jsonl logs afterwards, and have the dispatcher reply with only a count so its transcript does not flood the main context. Keep one folder per run with hard links to shared data files so repeats cannot see each other.