Bash alone reportedly scores higher than typed tools in enterprise-agent tests

According to DAIR.AI’s summary, a Microsoft paper compared five tool interfaces using Opus-4.8 and GPT-5.5. Bash alone scored 21.8–24.5 points higher than typed tools on TheAgentCompany and 4.8–7.4 points higher on APEX-Agents, while using 19%–72% fewer tokens. The summary says adding typed tools or agent-written tools on top of Bash gave no measurable gain. It attributes two recommendations to the authors: use Bash when execution can be sandboxed, and use programmatic tool calling—where the agent writes code against a fixed catalog—when compliance requires a fixed tool list.