LN: Tao et al. (2024) — MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution
Bibliographic Reference
Citation: Tao, W., Zhou, Y., Zhang, W., & Cheng, Y. (2024). MAGIS: LLM-based multi-agent framework for GitHub issue resolution. arXiv:2403.17927. https://arxiv.org/abs/2403.17927 (This paper is in the PUMA project knowledge PDFs as “MAGIS_LLM-Based_Multi-Agent_Framework_for_GitHub_Issue_ReSolution_2403.17927v2.pdf”)
Pass 1 — Bird’s Eye View (5 Cs)
| C | Assessment |
|---|---|
| Category | System proposal + empirical evaluation |
| Context | Addresses GitHub issue resolution with a team of specialised LLM agents |
| Correctness | Evaluated on SWE-bench. Improved over single-agent baselines. |
| Contributions | (1) Manager + Repository curator + Developer + QA team pattern; (2) Role-specific context retrieval; (3) Better than SWE-agent on SWE-bench subset |
| Clarity | Good. Clear role descriptions. |
Relevance: ⭐⭐⭐⭐
MAGIS’s GitHub issue team is directly analogous to PUMA’s multi-agent PM team. The “Repository curator” role (finds relevant files) maps to PUMA’s “historical issue retriever” in Stage 4.
PUMA Connection
MAGIS demonstrates that issue triage and resolution benefit from specialist agent decomposition. PUMA’s Stage 1 (triage classification) is a simpler version of MAGIS’s full issue resolution pipeline. Reference for positioning PUMA in the issue-handling landscape.