Bash alone beat typed tool catalogs by 21.8-24.5 points on TheAgentCompany while using up to 72% fewer tokens

arXiv 2609.11999·high signalXBlueskyLinkedInCopy link

A controlled study across five tool interfaces (typed tools, typed+bash, bash alone, bash with agent-synthesized persistent tools, and programmatic tool calling) on TheAgentCompany and APEX-Agents with Opus-4.8 and GPT-5.5 found bash alone won on both benchmarks: +21.8 to +24.5 pp on TheAgentCompany, +4.8 to +7.4 pp on APEX-Agents, with 19-72% fewer total tokens. Adding typed tools or persistent tool synthesis on top of bash produced no detectable pooled gain. Programmatic tool calling (programs restricted to a typed catalog) cut tokens versus direct typed calls but generally lost to bash on both quality and cost, so the authors recommend bash where arbitrary execution can be isolated and PTC only where a fixed catalog is a compliance requirement.