Dispatching Parallel Investigations

Think literature surveys across different subtopics, replicating the same analysis on separate datasets, or pre-specified robustness checks. The key constraint is intellectual honesty: it's built for cases where you'll report all results, not for fishing through 20 specifications to cherry-pick the one that works. Each agent gets a focused scope and exact instructions, runs its investigation independently, then you synthesize and verify everything. The documentation is unusually direct about the p-hacking risk, which is refreshing. Works well when parallelism actually saves time and you were going to run all the analyses anyway.