Ten agents, 500 files, one naming rule: the migration that did not need parallelism

The five instructions handed to each agent are worth reading item by item: find the legacy logging, replace it with the structured logger, use the correct channel, preserve the business logic, finish your assigned files ^[5]. By the author's own later accounting, most of that is not judgment. Finding legacy calls is a command that prints file and line number ^[10]. The channel is a pure function of the file path, produced by a generator and validated once into a map ^[11]. Knowing which files are done is a status field, which is what a JSON file, a database or a task queue exists to hold and what an LLM context is not built for ^[9]. Preserving business logic is genuine work, but it is local to one file and needs no coordination. Mapped against the mechanisms the author eventually assigns them, none of the five items needs a per-batch opinion ^[4]. One of them needs there to be exactly one opinion.

That one is the channel. Every script required its own ^[2]. Uniqueness is a property of the whole set, not of a batch, and an agent holding fifty files ^[4] can only check those fifty. Two agents can each be locally correct and jointly wrong, and that defect is invisible in any single diff, because it does not exist in a file. It exists in the collection. What the post describes is a model being asked to reconsider naming conventions and uniqueness for every file ^[11], which is the same authority instantiated ten times over, each copy unable to see what the others decided.

Workflow state fails the same way. Pending, processing, completed, failed, skipped ^[9] only mean something when one writer owns them. Ten contexts each maintaining their own copy produce ten partial views and no arbiter, and there is nobody to ask which view is current.

Worth saying plainly: the post reports no token counts, no wall-clock comparison and no defect rate from the ten-agent attempt [dx]. The case is structural, not measured. What survives the restructure is that the concurrency did not vanish, it moved. The preprocessing stage still builds an inventory, a set of log locations and a channel map ^[17], and those are deterministic jobs that can run alongside each other without disagreeing about anything. The logging engine was already written and reachable through a common include path before the migration began ^[3]. The remaining work was clerical, and clerical work at that volume wants a registry, not a committee.