ByteDance’s Software Engineering Research Lab released Trae Agent in 2025 as an open-source CLI coding agent that used a multi-model ensemble approach to achieve top performance on the SWE-bench Verified benchmark. By routing different subtasks to the most capable model — Claude 3.7 Sonnet for planning, Gemini 2.5 Pro for analysis, and o4-mini for implementation — the agent reached 75.2% on SWE-bench Verified, among the highest scores for any open-source tool at the time.
The work was presented at the ASE 2025 workshop and contributed to understanding of how multi-model orchestration could outperform single-model agents. The ensemble gained an estimated 5-8 percentage points over the best single-model configuration by dynamically selecting the model best suited to each phase of issue resolution.
Trae Agent (the research CLI) is distinct from Trae IDE (ByteDance’s commercial developer tool for daily use). Both demonstrated ByteDance’s investment in AI coding technology as Chinese AI labs significantly closed the gap with US counterparts in applied coding agent capabilities.