Trae Agent is a research-grade CLI coding agent from ByteDance’s Software Engineering Lab, built to push the boundary of automated software engineering. It achieved 75.2% on SWE-bench Verified using a multi-model ensemble combining Claude, Gemini, and o4-mini, placing it among the highest-scoring open-source tools on that benchmark. The project was presented at the ASE 2025 workshop and is distinct from Trae IDE, ByteDance’s separate commercial IDE product.
Key capabilities
Multi-model ensemble execution - Trae Agent coordinates across multiple LLM backends simultaneously, routing subtasks to the model best suited for each step. This ensemble approach is central to its SWE-bench performance and allows developers to mix providers including Claude, GPT-4o, and Gemini within a single agent run.
Full repository operations - The agent performs file reads and writes, multi-file edits, git operations, and terminal command execution autonomously. It can take an issue description and carry out the full cycle of reading code, writing a fix, running tests, and committing changes without manual intervention.
YAML-based configuration - Behavior is controlled through a YAML config file, making it straightforward to specify model preferences, API keys, and task parameters without modifying source code. Installation is via pip directly from the main branch.
Research-oriented extensibility - As an Apache 2.0 open-source project, Trae Agent is designed for experimentation and extension. Researchers and advanced developers can inspect, fork, and modify the agent loop, tool definitions, and model routing logic freely.
Autonomy level
Trae Agent operates at autonomy level 4, meaning it executes multi-step software engineering tasks end-to-end with minimal human checkpoints. Given a task description, it will independently read the codebase, plan a solution, implement changes across multiple files, run validation steps, and commit results. Human oversight is typically applied at task initiation and final review rather than during execution.
Strengths
- Highest SWE-bench Verified score (75.2%) among open-source coding agents at time of publication
- Multi-model ensemble architecture allows mixing Claude, Gemini, GPT-4o, and o4-mini for optimal results
- Apache 2.0 license permits unrestricted commercial and research use
- Pip-installable with YAML configuration, keeping setup friction low
- Backed by ByteDance SE Lab research with ASE 2025 workshop presentation
- Full source available for inspection, modification, and contribution
Limitations
- No formal versioned releases; installation is from the main branch, which may introduce instability
- Research-oriented design means it is not optimized for daily development workflows or IDE integration
- Achieving peak SWE-bench performance requires API keys for multiple LLM providers simultaneously
- Documentation and community support are thinner than commercial alternatives
- Not to be confused with Trae IDE, which is a separate ByteDance commercial product with different capabilities and licensing