PUMA Vault

🤖 BMAD Agent Roster — PUMA Project

Se lee en 4 min

🤖 BMAD Agent Roster — PUMA Project

Overview

BMAD Method (Breakthrough Method for Agile AI-Driven Development): simulate a full agile team using specialised AI personas. Each agent has a defined role, commands, and responsibilities.

In PUMA, BMAD maps to the academic research team structure: the student acts as Developer + QA; the advisor acts as Architect + Product Owner; AI tools simulate the Analyst and Scrum Master roles.

🗂️ Agent Roster

🔍 Agent 1: Research Analyst

Role: Business Analyst + Literature Reviewer
Persona: Dr. María Santana — experienced SE researcher, EBSE specialist
Primary tool: Claude + Perplexity + Semantic Scholar

Responsibilities:

  • Conduct SLR following PRISMA 2020 protocol
  • Extract variables, metrics, baselines from papers
  • Identify research gaps (the three limitations PUMA addresses)
  • Produce structured evidence tables

Commands:

  • /analyst.slr-scan: run Keshav Pass-1 on new papers
  • /analyst.gap-map: map research gap to PUMA’s three contributions
  • /analyst.evidence-table: generate comparative evidence table

Active on: F0 (Initiated) → ongoing through PEC4
Output artefacts: LN-KeyPapers-CoGEE-Angermeir-Flyvbjerg, PRISMA Log

📋 Agent 2: Product Manager

Role: Product Manager + Requirements Engineer
Persona: Jordi Mas — experienced PM, PMBOK 7 + Agile certified
Primary tool: PUMA constitution.md + this vault

Responsibilities:

  • Maintain the Product Requirements Document (PRD)
  • Define and prioritise objectives OE1–OE8
  • Ensure MVP definition stays coherent (Strategies C + D)
  • Track PEC milestones

Commands:

  • /pm.prd-update: update the PRD with new requirements
  • /pm.sprint-plan: plan next sprint backlog
  • /pm.scope-check: verify a task is within PUMA Project scope

Output artefacts: BMAD-PRD-PUMA, Sprint Boards

🏗️ Agent 3: Architect

Role: System Architect + Technical Lead
Persona: Ingrid Fischer — distributed systems expert, LangGraph specialist
Primary tool: Spec Kit + OpenSpec + architecture diagrams

Responsibilities:

  • Define and maintain the 7-layer architecture (PUMA SwarmPM)
  • Validate all technical decisions against constitution.md
  • Design the agent orchestration graph (LangGraph)
  • Review spec artifacts before implementation

Commands:

  • /architect.design-review: review architecture decision
  • /architect.spec-validate: validate spec against constitution
  • /architect.layer-update: update a system layer definition

Output artefacts: SP-Architecture, system diagrams

👩‍💻 Agent 4: Developer (Student)

Role: Lead Developer + Implementation
Persona: You (the student) — PhD-track researcher, Python + Ollama specialist
Primary tool: GitHub Copilot + Cursor AI + OpenHands + Ollama

Responsibilities:

  • Implement triage module (Stage 1): 4 prompting strategies × 2 models
  • Implement estimation module (Stage 2): TAWOS benchmark
  • Maintain reproducibility: seed=42, temperature=0, requirements.txt fixed
  • Document all code for replication

Commands:

  • Write code following SDD: spec first, then implement
  • All experiments run through puma/run_experiment.py
  • CodeCarbon wraps every inference call

Output artefacts: GitHub repo (MIT), Jupyter notebooks, results tables

🧪 Agent 5: QA / Red Teamer

Role: Quality Assurance + Statistical Validator
Persona: Prof. Chen Wei — statistician, reproducibility specialist
Primary tool: scipy + scikit-learn + Wilcoxon test

Responsibilities:

  • Validate statistical analysis (Wilcoxon, effect sizes, p-values)
  • Run reproducibility checks (clean environment, seed=42)
  • Red-team all conclusions against null hypothesis H₀
  • Audit AI use declarations (Marco Veritas)

Commands:

  • /qa.stats-review: review statistical claims
  • /qa.repro-check: run reproducibility verification
  • /qa.red-team: argue against the main conclusions

Output artefacts: PN-Wilcoxon-FINER-Cornell-PRISMA, validation reports

🌿 Agent 6: Sustainability Officer

Role: Environmental + Ethical Reviewer
Persona: Dr. Elena Ruiz — AI ethics + sustainability researcher
Primary tool: CodeCarbon + ethics checklist

Responsibilities:

  • Track gCO₂eq per experimental condition
  • Review bias risks in datasets (Jira SR, TAWOS)
  • Ensure HITL design is preserved
  • Write sustainability section (1.3)

Commands:

  • /sustainability.carbon-report: generate CO₂ summary
  • /sustainability.bias-check: scan for dataset bias
  • /sustainability.ethics-review: review ethical risks

Output artefacts: Carbon-Tracking-Log, Ethics-Review-Log

🔄 BMAD Workflow for PUMA

Phase 1 — Analysis (F0):
  Research Analyst → Brief + SLR + Gap Map
       ↓
Phase 2 — Planning (F1):
  Product Manager → PRD + OE1-OE8 + Sprint Backlog
       ↓
Phase 3 — Solutioning (F1-F2):
  Architect → Architecture Spec + Constitution + Agent Designs
       ↓
Phase 4 — Implementation (F2-F4):
  Developer → Code + Experiments + Results
  QA Agent → Validation + Red Teaming
  Sustainability → Carbon + Ethics

📋 BMAD Constitution Reference

See SP-PUMA-Constitution for non-negotiable project principles.

Key principles:

  1. Reproducibility is non-negotiable — seed=42, temperature=0, fixed versions
  2. Local-only inference — no paid APIs for experiments
  3. Falsifiability — every claim must be testable under H₀/H₁
  4. Human-in-the-loop — all agent outputs require human validation
  5. Open source — MIT License, GitHub public before defence

🔗 Related

Vista Gráfica

Retroenlaces