Prompt-side Playbook Transferability and Cost Scaling in Tool-Using Agents

Transferring distilled playbooks across benchmarks like ALFWorld and TAU2-Bench shows heterogeneous results, often requiring target-side tuning to remain effective. On XBench-DeepSearch, playbook transfer preserves heuristics but triggers query repetition and significant cost inflation due to context-runtime shifts.