Architecture
Benchmarked against: Anthropic — (No direct equivalent — SuperPortia reference) Scope: System architecture overview for the SuperPortia fleet
This page provides the high-level architecture of the SuperPortia multi-agent system.
Fleet architecture
Dual Agentic Lines
| Aspect | Line 1 | Line 2 |
|---|---|---|
| Stack | Claude Code + Agent SDK | LangGraph + LangChain |
| Engine | Claude family | Groq, Gemini, DeepSeek, Mistral, Zhipu |
| Role | Architecture, decisions, code | Research, classification, batch ops |
| Cost | $$$$ | Cents or free |
Universal Brain architecture
Work Order system
| Feature | Implementation |
|---|---|
| State machine | 8 states with enforced transitions |
| RBAC | Captain/PM/Assignee/Requester roles |
| Transition gates | Mandatory fields at specific transitions |
| Audit trail | Full transition history per WO |
| ID format | WO-YYYY-MMDD-NNN (auto-generated) |
MCP server topology
Agent configuration stack
Data flow
Key architectural decisions
| Decision | Rationale |
|---|---|
| Cloud UB as single brain | Fleet-wide consistency, any agent on any ship |
| Dual agentic lines | Mutual rescue, cost optimization |
| MCP for tool access | Standard protocol, agent-native discovery |
| WO for task management | Auditable, RBAC, state machine enforcement |
| Progressive Disclosure | Token efficiency, load on demand |
| English-only UB | Embedding quality, cross-agent consistency |
| Hooks for enforcement | Tool-enforced rules, not discipline-based |
Related pages
| Page | Relationship |
|---|---|
| Agent SDK Overview | Dual agentic lines |
| Ingest Fragment API | UB ingestion pipeline |
| WO Lifecycle | Work order system |
| MCP Tool Discovery | Tool topology |