Recoleta
GitHub
Source note

TraceDev: A Traceability-Driven Multi-agent Framework for Requirement-to-Code Development

Authors
Mingyu Chen, Yakun Zhang, Zihao Xie, Yixing Luo, Jinrui Xu, Cuiyun Gao, et al.
Source
arXiv
Published
Collection
Software Intelligence
Software Foundation ModelCode IntelligenceAutomated Software ProductionMulti Agent Software EngineeringRequirements Traceability
Open arXivMarkdown source

Summary

TraceDev is a five-agent framework that generates repository-level code from complex, multi-step use cases. Its central mechanism is a heterogeneous traceability graph linking requirements, design models, and code so agents can detect omissions and refine artifacts.

Problem

  • Existing requirements-to-code systems often use simplified, single-sentence tasks that do not represent use cases with multiple functional points and semantic constraints.
  • They generally lack explicit traceability between requirements, designs, and implementations, making omissions and cross-stage semantic errors difficult to detect.
  • This matters because incomplete or incorrect repository-level implementations reduce functional correctness, maintainability, and confidence in automated software production.

Approach

  • TraceDev uses five role-specific LLM agents: Requirement Refiner, Designer, Developer, Tester, and Validator.
  • The Requirement Refiner normalizes use-case terminology, infers missing subjects, rewrites ambiguous steps, and applies iterative syntactic verification.
  • The Designer creates PlantUML class and sequence diagrams; the Developer generates code from those designs; and the Tester generates and executes tests, sending failures back for bounded self-correction.
  • The Validator builds a heterogeneous directed graph containing requirement entities, design elements, and code files. LLM semantic matching links requirements to designs, while AST-based matching links designs to code and exposes missing implementations.

Results

  • Across the ETOUR and SMOS datasets, the evaluation covers 125 use cases and compares TraceDev with ChatDev and MetaGPT.
  • On ETOUR, TraceDev reports a 71.72% semantic-coverage rate, exceeding ChatDev by 51.66% and MetaGPT by 75.14%.
  • On ETOUR, TraceDev achieves a 53.63% success rate, exceeding ChatDev by 129.19% and MetaGPT by 186.64%.
  • On SMOS, TraceDev achieves a 56.82% success rate, surpassing the baselines by up to 340.80%.
  • The excerpt does not provide ablation results, statistical significance, execution cost, or complete per-dataset metrics, so it supports improved benchmark performance but not a separate causal measurement of the traceability graph's contribution.

Link