ByteDance Trae Agent: Ensemble Voting as Test-Time Scaling for Repository-Level Bug Fixing
> SWE-bench Verified 75.20% Pass@1 / 14 authors / MIT license / a repo dormant for 14 months that still matters — this marks a key inflection point in AI coding for 2026.
The paper and its claim
ByteDance's paper *Trae Agent: An LLM-based Agent for Software Engineering with Test-time Scaling* (arXiv 2507.23370), published July 31, 2025, with Pengfei Gao and Zhao Tian as equal-contribution first authors and 14 authors total, framed repository-level issue fixing as an optimization search problem: find, in the largest possible solution space, a patch that passes all golden tests. At release it ranked first on the SWE-bench Verified public leaderboard with 75.20% Pass@1.
Test-time scaling — the idea that one model, sampled multiple times, yields better results — had a catch: multiple attempts must be synthesized coherently. At repository level, the model must also understand dependencies, affected files, and which tests to run.
The three-stage workflow
1. Patch generation — semantic retrieval first, then a coder agent produces N structurally diverse drafts (via different temperatures/seeds) to maximize ensemble diversity. 2. Patch pruning — deduplication merges near-identical patches (naive multi-sampling often yields mostly duplicate patches); regression testing runs the repo's non-target tests only when candidates remain too many, avoiding overhead on simple fixes. 3. Patch selection — a selector agent performs static review (reading referenced, modified, and dependency-connected files) and dynamic verification (running generated tests, collecting execution traces), then selects via majority voting across both channels.
Results
- Models tested: Gemini 2.5 Pro, Claude 3.7 Sonnet, GPT-4.1 on SWE-bench Verified, against four SOTA ensemble baselines (Augment, Augment w/ Pruning, DeiBase, DeiBase w/ Pruning).
- Average Pass@1 improvement of 10.22%, given in the paper as a 5.83%–14.60% range depending on baseline (largest under DeiBase, smallest under Augment).
- Headline result: 75.20% Pass@1 on SWE-bench Verified — leaderboard first as of July 2025.
- Full module decoupling: CLI layer, agent decision layer, tool layer, and model adapter layer communicate through unified interfaces, making ablation studies cheap (stub out a module and run).
- Six provider adapters: OpenAI, Anthropic, Google Gemini, OpenRouter, Doubao, and Ollama in one codebase — rare breadth in 2025 agent repos.
- YAML configuration with environment-variable fallback, allowing model/ensemble/prompt changes without code edits.
- arxiv.org/abs/2507.23370
- github.com/bytedance/trae-agent
- agentlist.top/en/projects/bytedance-trae-agent
- startuproast.com SWE-bench leaderboard compilation (2026-09)
Note on secondary reports: figures like 32.7% (SWE-bench Verified) / 43.8% (Lite) circulating in secondary coverage reflect single-model baselines, not the ensemble Pass@1 of 75.20%; the two are not directly comparable.
Why it was a research-oriented open-source release
The README positions Trae Agent as "designed for researchers", backed by concrete choices:
Quick start: git clone https://github.com/bytedance/trae-agent.git && cd trae-agent && uv sync --all-extras && source .venv/bin/activate && trae-cli run "..." (Python 3.12+, uv recommended).
Positioning and legacy
Trae Agent was the first agent-based ensemble reasoning approach for repository-level issue resolution, upgrading ensemble from a sampling trick to a system component. By September 2026 the repo shows 12.1k+ stars but has been dormant for over seven months — it stayed on July 2025 model baselines while the leaderboard moved on (Claude Sonnet 4.6 and Gemini 3.6 Flash around 79.6%).
Two legacies remain: (1) the four-layer decoupled engineering pattern, adopted by many 2026 open-source agents; (2) the generate-prune-select "Trae-style" ensemble workflow, repeatedly cited (with variations like verifier-based filtering or LLM-as-judge) in 2025–2026 agent literature.
The core idea — letting multiple agents explore repository-level tasks independently and then converge — is now embedded in Claude Code, Codex, Cursor, Trae IDE, and Kiro. The bird stopped flying, but the idea it carried now flies in every modern issue-fixing agent.