Six agents editing one header is not a strategy
Parallelism is useful when the work separates cleanly. It gets messy when the split is just “give another agent something to do”.
Retrospective field note for 2025-10-13. Reconstructed in November 2026 from dated sources, not originally published on the archive date.
Parallelism is useful when the work separates cleanly. It gets messy when the split is just “give another agent something to do”.
The discussion around the Agent SDK makes this worth being specific about. Independent source research can be divided. Three workers changing the same navigation component need a merge plan before they need encouragement.
Assign an output and an owner. Name the files or decisions each worker controls. Keep the shared constraints in one place, and pass corrections to everyone affected. A new user instruction should not arrive at integration as a surprise.
Leave time for that integration. A content format can be correct on its own and break the page generator. A layout fix can pass its tests while another change invalidates the screenshot. Someone has to check the combined result.
Measure the time until that combined result is verified. Include briefing, waiting, merging and repair. Counting only the minutes agents were generating in parallel produces a nice acceleration number and tells the person waiting very little.
I’d rather have two well-scoped workers and a clean handoff than six producing overlapping patches. More capacity helps once the work has boundaries. Before that, it mainly makes the confusion arrive faster.