The paper studies the cost-quality trade-off when switching between low-cost (LC) and high-cost (HC) during a task, a process called handoff. Using pairs of Claude and GPT models, the authors vary handoff direction, timing, and the amount of trajectory information passed to the receiver. They find that escalating from LC to HC with full trajectory recovers less than half the quality gap while significantly increasing cost, a penalty they term the handoff tax. By contrast, downshifting from HC to LC yields a favorable cost-quality point.
Switching Models Mid-Task Imposes a Cost and Quality Penalty
Experiments with Claude and GPT model pairs show that escalation often fails to recover stronger-model performance, while downshifting provides a better cost-quality trade-off.
Industry
Roy Ganz · Mor Shpigel Nacson · Adi Kalyanpur · Ron Litman
AWS
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
Handoffs between cheap and expensive coding agents incur a quality-cost penalty, the handoff tax, which varies with direction and trajectory information.
Before you read this
Two ideas this paper builds onRead it as
Ganz et al. tested coding agents that switched between lower-cost, lower-capability models and higher-cost, higher-capability models during long-running tasks.
The authors paired lower-cost, lower-capability and higher-cost, higher-capability models from the Claude and GPT families.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.