Commerce Agent Bench tops out at 60.7% on 107 tasks

Alibaba International’s Accio team has open-sourced Commerce Agent Bench, a 107-task suite spanning browser, CLI, file, and API/MCP workflows; its published Pi table tops out at 65 passes out of 107 (60.7%) for Claude Opus 5, according to the benchmark repository. Each run must change state in a fresh container and preserve its trajectory, but the published table is a snapshot and uses Gemini 3.1 Pro as judge. Teams piloting long-horizon agents should add state-changing commerce tasks to their bake-offs and compare harness, verifier, and model together, extending the archive’s evidence that harness choice can swing same-model coding outcomes.