Step Function has made its terminal programming agent, Step Code, available as open-source software. Licensed under the MIT License, Step Code facilitates a comprehensive workflow encompassing coding, debugging, execution, and deployment—all within the terminal environment. In the Terminal Bench 2.1 assessment, Step Code demonstrated outstanding performance, achieving a tie for first place with an impressive 80.9% pass rate while utilizing fewer Tokens. Furthermore, in the self-constructed long-term benchmark Multi-Frame evaluation, Step Code excelled once again, securing the top position among six Agent Harnesses with a 73.3% pass rate. Notably, it maintained the lowest average Token consumption at 5.09M, underscoring its efficiency.
