Decomposition Buys Integrity, Not Yield
Splitting work across agents preserves focus and lowers context exposure, but it reduces how much discovered material reaches the root.
Rong He models decomposition as a tree and finds that deeper agent structures do not improve yield; under the measured scaling, flat architectures are best for passing findings upward. Production traces estimate a per-tier penalty, with alignment losses compounding the drop. The paper argues depth is still useful when the value is integrity, smaller root context, or cost control, not maximum recall. In the measured data, only a small slice of sessions appear worth delegating, and delegation is used early rather than as a response to a filling context. ArXiv · AI/CL/LG's note
Rong He models decomposition as a tree and finds that deeper agent structures do not improve yield; under the measured scaling, flat architectures are best for passing findings upward. Production traces estimate a per-tier penalty, with alignment losses compounding the drop. The paper argues depth is still useful when the value is integrity, smaller root context, or cost control, not maximum recall. In the measured data, only a small slice of sessions appear worth delegating, and delegation is used early rather than as a response to a filling context. ArXiv · AI/CL/LG's note
score 4