Decomposition Buys Integrity, Not Yield

Problem

This work addresses a gap in understanding the impact of task decomposition in multi-agent systems on information discovery. The authors highlight that existing literature does not adequately explore how decomposing tasks affects the integrity of findings produced by agents. This paper is a preprint and has not undergone peer review.

Method

The authors model task decomposition as a tree structure where an agent retains items based on a probability function, denoted as $r(b)$. When $r(b) = 1/b$, each tree structure is capable of delivering one finding regardless of the task size or shape. The model's verification achieved a precision of $2.4 imes 10^{-15}$ across 20,000 random irregular trees. An alternative probability function is introduced as $r(b) = Cb^{-δ}$, where a depth-$k$ tree yields findings proportional to $C^k N^{1-δ}$. The authors utilized 600 production deep-research traces to estimate $δ$ at 0.34, with a confidence interval of [0.30, 0.38]. Observations at a hop with $b=1$ occurred 550 times, yielding an observed constant $C = 0.571$ with a confidence interval of [0.527, 0.615] across 16,082 hops. The alignment cost was analyzed using 1,012 annotated multi-agent traces, resulting in a mean alignment cost of $μ= 0.939$ and a per-tier penalty of $Cμ= 0.536$. The cost analysis indicates that production flat agents bill as $N^{1.39}$, with two tiers becoming more efficient than flat agents at 403 findings.

Results

The verification metric achieved was $2.4 imes 10^{-15}$, with no comparative baselines reported. The finding yield was quantified with $δ= 0.34$ [0.30, 0.38], and the observed constant $C$ was $0.571$ [0.527, 0.615]. The authors found that only 0.7% to 11.3% of production sessions are worth delegating, compared to 7.8% that currently do.

Limitations

The authors note that depth cuts exposure reduces the number of items from $N$ to $N^{1/k}$. Additionally, the delegation process does not account for a filling context, which may limit its applicability in certain scenarios.

Why it matters

This research has significant implications for the design and optimization of multi-agent systems, particularly in enhancing information discovery through effective task decomposition. Understanding the trade-offs between integrity and yield can inform future developments in agent-based systems and their applications in complex environments.