LN: Arora et al. (2024) — MASAI: Modular Architecture for Software-Engineering AI Agents

Bibliographic Reference

Citation: Arora, D., Sonwane, A., Wadhwa, N., et al. (2024). MASAI: Modular architecture for software-engineering AI agents. arXiv:2406.11638. https://arxiv.org/abs/2406.11638 Affiliation: Microsoft Research India

Important Note

Overview

The bibliography entry attributes this paper to “Xie et al.” — this is incorrect. The verified first author is Daman Arora.


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategorySystem proposal + empirical evaluation
ContextBuilds on SWE-bench; targets GitHub issue resolution with modular agents
CorrectnessAchieves 28.33% on SWE-bench Lite (300 GitHub issues). State-of-the-art at publication. Clear ablations.
Contributions(1) Modular sub-agents with well-defined objectives; (2) Information gathering across repo; (3) Avoiding long trajectories; (4) SOTA on SWE-bench Lite
ClarityExcellent. Clear architecture diagrams.

Relevance: ⭐⭐⭐⭐

MASAI’s modular sub-agent architecture is directly applicable to PUMA: separate agents for triage, estimation, and scheduling, each with well-defined scope.


Pass 2 — Key Points

MASAI decomposes software engineering into sub-problems, each handled by a specialist agent. The key innovation: each sub-agent has a narrow objective and specific tools, reducing error accumulation. This is the opposite of single-agent approaches that try to do everything in one context.

PUMA application:

  • Triage sub-agent: classify priority of one issue at a time
  • Estimation sub-agent: predict story points for one user story
  • These could be orchestrated by a Manager Agent (MASAI’s orchestrator pattern)

PUMA Integration

MOCs