Agent SDK Overview
Benchmarked against: Anthropic — Agent SDK Overview Architecture: Dual Agentic Lines — Claude (30%) + LangGraph (70%) Principle: Both lines are truly agentic — autonomous perceive-reason-act-learn systems
SuperPortia operates two parallel agentic lines. This is not "pick one" — both run simultaneously with mutual disaster recovery.
Dual Agentic Lines
Line comparison
| Aspect | Line 1: Claude Agentic | Line 2: LangGraph |
|---|---|---|
| Share | ~30% of operations | ~70% of operations |
| Stack | Claude Code + Agent SDK | LangGraph + LangChain + MCP Adapters |
| Role | Architecture, decisions, code ops | Research, analysis, classification, patrol |
| Engine | Claude Opus / Sonnet / Haiku | Groq / Gemini / DeepSeek / Mistral / Zhipu |
| Tool access | All MCP servers + built-in tools | MCP via langchain-mcp-adapters |
| State | Conversation context + MEMORY.md | LangGraph checkpoints + UB |
| Cost | $$$$ (premium) | Cents (low-cost or free) |
| Best for | Architecture, key decisions, delegation | Batch processing, research, classification |
Mutual rescue
The dual-line architecture ensures resilience:
Captain's principle: "One goes down, the other can rescue." Neither line is disposable — both must be production-ready at all times.
Agent Anatomy — 13 Components
Line 1 designed the blueprint (Agent Anatomy 13 Components), and Line 2 implements it:
| # | Component | Description |
|---|---|---|
| 1 | Identity | Agent name, role, ship assignment |
| 2 | Memory | Short-term (context) + long-term (UB) |
| 3 | Perception | Input processing (text, images, files) |
| 4 | Reasoning | Decision-making logic |
| 5 | Planning | Task decomposition and sequencing |
| 6 | Action | Tool execution capabilities |
| 7 | Learning | Pattern recognition and correction capture |
| 8 | Communication | Inter-agent messaging (mailbox) |
| 9 | Self-monitoring | Heartbeat, health checks |
| 10 | Governance | Constitution, cost awareness, HITL |
| 11 | Context | Session management, handoff protocols |
| 12 | Tools | MCP server connections |
| 13 | State | Persistent state across sessions |
Causal relationship: Line 1 blueprint → Line 2 implementation. Line 1 is the pioneer, not the laggard.
Line 2 ecosystem
The LangGraph line uses these verified packages:
| Package | Version | Purpose |
|---|---|---|
langgraph | 1.0.8 | Agent graph framework |
langgraph-checkpoint-sqlite | 3.0.3 | State persistence |
langchain-mcp-adapters | 0.2.1 | MCP tool bridge |
langchain-google-genai | Latest | Gemini integration |
langchain-groq | Latest | Groq integration |
Choosing between lines
| Scenario | Recommended line |
|---|---|
| Architecture design, key decisions | Line 1 (Claude) |
| Code editing, file operations | Line 1 (Claude) |
| Research and intel gathering | Line 2 (Groq/Gemini) |
| Content classification | Line 2 (MTAAA pipeline) |
| Batch data processing | Line 2 (Groq = free) |
| Strategy analysis | Line 1 (Claude Chat) |
| Emergency quota conservation | Line 2 (all engines) |
Non-dogmatic component sharing
Lines can cherry-pick components from each other without vendor lock-in:
- Line 2 can use Claude as an engine for complex tasks (via
dispatch_work_orderwithengine: "claude") - Line 1 can delegate to Groq/Gemini for cost optimization (via
search_web,call_model) - Both share the same Cloud UB, WO system, and MCP infrastructure
The goal is agentic capability, not framework purity.
Related pages
| Page | Relationship |
|---|---|
| Agent SDK Quickstart | Get started with either line |
| Python SDK (LangGraph) | Line 2 deep dive |
| TypeScript SDK | Line 1 SDK reference |
| Line Migration | Moving between lines |