Trae Agent

About

Trae Agent exists to make agent research practical: production harnesses obscure their internals, which prevents controlled comparison of prompts, tools, and orchestration strategies. It provides a modular Python CLI in which a natural-language engineering task is executed through a small tool set — bash, str_replace-based file editing, sequential thinking — with interactive mode for iterative work and Docker mode for reproducible, isolated execution from an image, existing container, Dockerfile, or archive. Every run emits a trajectory log capturing LLM calls, tool usage, and step-by-step state, and Lakeview summarizes those steps for analysis; YAML configuration with environment-variable overrides keeps experiments scriptable. Agent researchers and engineers studying harness design are its primary users, and an accompanying arXiv technical report documents the architecture.

Other entries by bytedance

• DeerFlow — agent