One agent until it hurts

The temptation to split things up shows up early, and almost always too early.

Multi-agent setups look more sophisticated. Each specialist does what they do best, the work flows, the diagram looks great. In practice, every new agent adds a call, latency, an exchange format, and a point of failure, and none of that shows up in the diagram.

The rule is counterintuitive: one agent until it hurts. Split only when a single agent has already failed in a measured way, not when you imagined it would fail.

When it is time to split, the criterion is context, not the org chart. A good reason to create a subagent is that a subtask generates a lot of intermediate material the main agent does not need to carry, such as 40 pages of history that become five lines of summary. A bad reason is mirroring the company’s departments as agents.

In correspondence, three agents pay for themselves: one extracts facts, one writes, one checks against the source. If it were 40 letters a year instead of 4,000, one agent would handle it, and the split would be pure cost.