A new paper puts a number on the multi-agent hype: splitting a task across a tree of agents does not get you more of what those agents actually find.
The researchers model each agent as being handed a batch of b items and keeping one with probability r(b). If r(b) equals 1/b, every tree design returns exactly one finding no matter how many agents you add or how you arrange them. The authors verified this to a precision of 2.4 x 10^-15 across 20,000 random tree shapes. Real production systems fare a bit better than that worst case: across 600 deep-research traces, the yield decay exponent measured 0.34, and a dataset of 16,082 tool-call hops pinned the per-level retention rate at 0.571. Alignment adds its own tax, too. On 1,012 annotated multi-agent traces, about one brief in sixteen drifted off-target.
None of that makes decomposition pointless. Deeper trees still win on two fronts that have nothing to do with yield. They shrink what the root has to hold in context from N items down to roughly the k-th root of N, and they are cheaper to run since flat agents bill as N^1.39 rather than the N^2 an ever-growing context would predict. At equal spend, a two-tier setup overtakes a flat one once you pass 403 findings.
The sharpest number here might be this one. The model itself estimates that only 0.7 to 11.3 percent of production sessions are worth delegating to subagents, yet teams delegate 7.8 percent of them anyway, and delegation decisions do not even wait for context to fill up. They happen as an opening move. Call it evidence-based skepticism for anyone building agent orgs chart-first.