When to Delegate: A Measurement Study of Bounded Agent Teams for Personal Super Agents

Abstract Personal super agents coordinate many tools while remaining fast, safe, and accountable to one user. A popular design is a Router delegating to Workers. Recent work shows coordination can cost more than it returns, but transfer to personal agents—single-user context, side-effecting tools, cross-turn references, and refusals—is not established. We present the bounded-delegation runtime protocol, instantiate it in Eva, a 22-tool in-vehicle agent, and introduce Eva-Bench, 80 rule-scored tasks in ten categories. With one real model and tool set, we compare monolithic ReAct, bounded team, and unbounded team, and ablate budgets, reasoning, concurrency, guards, ledger, and injected failures. Delegation is not free: the main team pays 2.75×the latency and 2.59×the tokens. In one full-suite run it leads by one task (92.5% versus 91.2%), but three full-suite repetitions give 92.9% versus 93.3% (difference -0.4pp). Four post hoc temperature- matched runs average 90.6% (range 86.2–93.8%), so temperature alone does not explain the team’s small, unstable lead. We do not claim a general team advantage. The stable finding is the tax; service workflows and cross-turn references favor the team in repeated data, while safety and aggregate success do not. Bounded budgets reduce tail latency, reasoning is a net cost, and guards/ledger are null at oracle success. A selective-delegation policy reaches 95.0% at 6.27s in one run, versus 91.2% at 3.22s monolithic and 92.5% at 8.85s for the fixed team; it is exploratory. Delegation should be earned by request structure, and evaluation must report organizational overhead separately from task success.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11710191/v1
Primary Topic
AI-based Problem Solving and Planning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When to Delegate: A Measurement Study of Bounded Agent Teams for Personal Super Agents

Jipu Su
Research Square
AI-based Problem Solving and Planning
preprint

When to Delegate: A Measurement Study of Bounded Agent Teams for Personal Super Agents

Jipu Su
preprint en

Abstract

Abstract Personal super agents coordinate many tools while remaining fast, safe, and accountable to one user. A popular design is a Router delegating to Workers. Recent work shows coordination can cost more than it returns, but transfer to personal agents—single-user context, side-effecting tools, cross-turn references, and refusals—is not established. We present the bounded-delegation runtime protocol, instantiate it in Eva, a 22-tool in-vehicle agent, and introduce Eva-Bench, 80 rule-scored tasks in ten categories. With one real model and tool set, we compare monolithic ReAct, bounded team, and unbounded team, and ablate budgets, reasoning, concurrency, guards, ledger, and injected failures. Delegation is not free: the main team pays 2.75×the latency and 2.59×the tokens. In one full-suite run it leads by one task (92.5% versus 91.2%), but three full-suite repetitions give 92.9% versus 93.3% (difference -0.4pp). Four post hoc temperature- matched runs average 90.6% (range 86.2–93.8%), so temperature alone does not explain the team’s small, unstable lead. We do not claim a general team advantage. The stable finding is the tax; service workflows and cross-turn references favor the team in repeated data, while safety and aggregate success do not. Bounded budgets reduce tail latency, reasoning is a net cost, and guards/ledger are null at oracle success. A selective-delegation policy reaches 95.0% at 6.27s in one run, versus 91.2% at 3.22s monolithic and 92.5% at 8.85s for the fixed team; it is exploratory. Delegation should be earned by request structure, and evaluation must report organizational overhead separately from task success.

Research Square
City University of Hong Kong (HK)
AI-based Problem Solving and Planning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon