LN: Spichkova (2025) — Cognitive Agents for Agile PM

Bibliographic Reference

Citation: Spichkova, M., Georgievski, I., & Čizmić, B. (2025). Cognitive agents for Agile software project management. In Proceedings of EASE 2025 [Preprint]. arXiv:2508.16678.


Pass 1 Summary (5 Cs)

CAssessment
CategoryEmpirical evaluation + system proposal
ContextBuilds on CoGEE; related to PM+LLM emerging body of work
CorrectnessUses proprietary models (GPT-4) — limited reproducibility
ContributionsEvaluates 5 Agile PM tasks with LLM agents. Shows CoT helps for structured tasks.
ClarityGood. Task definitions clear.

Relevance: ⭐⭐⭐⭐⭐ (5/5)

Direct competitor / predecessor to PUMA


Pass 2 Key Points

Five PM tasks evaluated:

  1. Story point estimation
  2. Issue triage (priority classification)
  3. Sprint planning
  4. Risk identification
  5. Status reporting

Key limitation (relevant to PUMA): Uses GPT-4 via API. Not reproducible without payment. No systematic prompting strategy comparison. No carbon measurement.

PUMA’s differentiation:

  • Local models (no API cost)
  • Systematic 4-strategy prompting comparison
  • CodeCarbon measurement
  • Open-source reproducible benchmark

Metrics used: F1 (triage), MAE (estimation), BLEU/ROUGE (text tasks). PUMA uses same for consistency.


PUMA Integration

Used in: Section 1.1 (gap: reproducibility + systematic prompting), SLR evidence table

Supports: Research gap justification (Limitation 2: no systematic prompting strategy comparison)


Pass 3 TODO

  • Reconstruct their triage prompting approach
  • Compare their F1 results to PUMA baselines when available
  • Generate permanent note: insight about task-specific prompt design

🔗 Connected Notes

Superseded by: LN-Cinkusz-2025-CognitiveAgentsAgilePM (corrected author metadata — Cinkusz is first author)

Permanent notes:

PUMA project:

MOCs: