🏢 Smart PMO Vision — PUMA Stage 5
Overview
This is aspirational scope — beyond the MVP. Documents the long-term vision informed by the PEC1 correction feedback. Serves as the “north star” for future work section of the PUMA Project.
What is a Smart PMO?
A Smart PMO (Project Management Office) is a PMO augmented by AI agents that:
- Automatically triage incoming issues and work items
- Estimate effort with calibrated confidence intervals
- Detect risk patterns from historical data (Reference Class Forecasting)
- Generate reports and status summaries
- Escalate anomalies to human managers
- All under explicit human-in-the-loop governance
The PUMA SwarmPM Architecture (7 Layers)
Based on the PEC1 correction document:
| Layer | Name | Technology | Status |
|---|---|---|---|
| 1 | Ingestion (Sensors) | Jira API, GitHub webhooks, LlamaIndex | Stage 4+ |
| 2 | Context & Memory (RAG) | Redis (short-term), Qdrant (long-term vectors) | Stage 4 |
| 3 | Reasoning (Inference) | Ollama + Llama 3.2 + Mistral | Stage 1–3 ✅ |
| 4 | Orchestration (Brain) | LangGraph cyclic state graph | Stage 4–5 |
| 5 | Execution (Actuators) | OpenHands, Warp AI Terminal | Stage 5 |
| 6 | Governance (Control) | Pydantic AI validation, OPA policies | Stage 3+ |
| 7 | Interface (Dashboard) | React + WebSockets, human approval gates | Stage 5 |
Agent Specialisation (BMAD → SwarmPM)
| BMAD Role | SwarmPM Agent | Specialisation |
|---|---|---|
| Analyst | Research Analyst | SLR, gap mapping, evidence extraction |
| Product Manager | Lead Orchestrator | Issue routing, sprint planning coordination |
| Architect | Risk Analyst | Monte Carlo simulation, CPM, risk flags |
| Developer | Triage Agent | Issue classification (Stage 1 MVP) |
| QA | Estimation Agent | Story point estimation (Stage 2) |
| Sustainability | Carbon Monitor | CodeCarbon integration, sustainability reporting |
Why Swarm Over Single Agent?
Three scientific justifications (from PEC1 correction):
-
Domain specialisation: a single general LLM has systematic bias when acting as expert in telemetry, people management, and architecture simultaneously. Compartmentalisation reduces accumulated error.
-
Multi-agent deliberation: agents can debate. The Risk Analyst can challenge the Estimation Agent’s prediction if it detects decreasing team velocity trends.
-
Resilience: swarm architecture has no single point of failure. If one agent fails, the orchestrator reroutes or requests human intervention.
Research Questions for Stage 5
- Does deliberation between agents improve triage/estimation quality vs. single agent?
- What governance policies (OPA rules) are necessary for safe autonomous PM assistance?
- What is the gCO₂eq cost of swarm coordination overhead vs. accuracy gain?
Connection to PUMA Project
The Smart PMO vision is documented in:
- Section 4: Conclusiones y trabajos futuros — as “Líneas de investigación futura”
- Section 1.2: Objetivos — Etapa 5 (🔴 Opcional)
- BMAD-PRD-PUMA — Non-requirements for MVP
References
- PN-KeyConcepts-Agents-Reproducibility-RedTeam — Agent OS, HITL
- PN-RAG-Embeddings-VectorDB — Layer 2 memory/context
- PN-MultiAgent-ArchitecturePatterns — Specialisation rationale
- PN-Agentic-Science-Paradigm — Closed-loop science parallel
- PN-ReAct-AgentPattern — Agent reasoning loop
- PER-Flyvbjerg-Bent — Reference Class Forecasting
- BMAD-Agent-Roster — Current BMAD team
- SP-Architecture — Stage 1–3 architecture
- PR-PUMA-Ch5-Discussion — Future work section
- MOC-PUMA-Master