Developer automates spec-driven development, achieving 94% on Terminal Bench 2.1.

The workflow begins with a specification document that guides a language model before task execution. To streamline this process, the developer built a supervisor agent that delegates design and implementation tasks to worker subagents. Using Codex’s app server framework and the GPT-5.6 Sol model, the supervisor sizes the task, requests a design document, receives the document, and then instructs the worker to produce an implementation spec and the corresponding code.

The system was evaluated on Terminal Bench 2.1, where it completed 94% of the required tasks. The developer reported that the automation reduced repetitive steps and improved efficiency for feature development, debugging, and greenfield projects. The approach relies on file reading and agent invocation rather than constructing a dedicated coding harness, leveraging existing solutions such as Pi, OpenCode, and Codex app server. The experiment demonstrates the feasibility of automating spec-driven development with current language model tools.