Skip to main content

Line Migration Guide

Benchmarked against: Anthropic — Migration Guide Architecture: Dual Agentic Lines Principle: Not replacement — mutual rescue and cross-learning

Migration between Line 1 (Claude Agentic) and Line 2 (LangGraph) is not about abandoning one line. It's about moving specific workloads to the most appropriate line while maintaining both.


Migration directions

DirectionWhenExample
Line 1 → Line 2Task proven on Claude, ready for low-cost automationResearch patrol, batch processing
Line 2 → Line 1Task too complex for LangGraph, needs Claude reasoningArchitecture decisions, novel problems
Line 1 blueprint → Line 2 implementationDesign in Claude, build in LangGraphAgent Anatomy 13 → superportia_agent.py

What migrates

AssetMigrates?How
Knowledge (UB)SharedBoth lines access Cloud UB — no migration needed
Work ordersSharedBoth lines use WO system
Agent messagesSharedBoth lines use mailbox
Tool definitionsNeeds adaptationMCP tools → LangGraph tool bindings
Prompts/RulesNeeds adaptationCLAUDE.md → LangGraph system prompts
State managementDifferentContext window → Checkpoint persistence

Migrating a task from Line 1 to Line 2

Prerequisites

  • Task is well-defined and repeatable
  • Task doesn't require Claude-only capabilities (file ops, complex reasoning)
  • A suitable low-cost engine exists for the task

Steps

  1. Document the Claude workflow — What tools, what sequence, what decisions
  2. Map to LangGraph graph — Nodes for each step, edges for flow
  3. Select engine — Groq (free), Gemini (cheap), or other
  4. Build and test — Create LangGraph agent, verify output quality
  5. Compare — Run same task on both lines, compare results
  6. Switch — Route future WOs to Line 2 dispatch

Example: Research patrol

Line 1 (before): Opus agent searches web, analyzes, ingests
Cost: ~$2-5 per patrol

Line 2 (after): LangGraph + Groq searches, summarizes, ingests
Cost: $0 (free Groq)

Migrating a task from Line 2 to Line 1

When needed

  • LangGraph agent produces poor quality output
  • Task requires file operations (only Claude can do this)
  • Task needs complex multi-step reasoning
  • Engine limitations (tool calling, hallucination)

Steps

  1. Identify failure mode — Why Line 2 isn't working
  2. Create WO with engine=claude — Route to Claude dispatch
  3. Or use Claude Code directly — Interactive session

Shared infrastructure advantage

The key insight: both lines share Cloud UB, WO system, and messaging. This means:

  • No data migration — Same UB, same WOs
  • Gradual migration — Move one task at a time
  • Easy rollback — Switch back by changing dispatch engine
  • Cross-learning — Line 1 insights improve Line 2, and vice versa

Anti-patterns

Anti-patternWhy it fails
Migrating everything at onceOne line fails, no fallback
Abandoning a lineLoses mutual rescue capability
Using Line 2 for architectureLow-cost engines lack deep reasoning
Using Line 1 for batch workExpensive, wastes Opus tokens

PageRelationship
Agent SDK OverviewDual lines architecture
Choosing an EngineEngine selection
Dispatch ModesHow dispatch routes to engines