1. This Prototype Was Built Using Claude (Anthropic)
This UI, all code snippets, and the entire interaction logic were generated using Claude Sonnet 4.5 by Anthropic. No LangGraph, LangChain, RAG pipeline, or vector database is active in this file — it is a self-contained HTML/CSS/JS prototype.
2. Production System Stack
| Component | Technology | Purpose |
|---|---|---|
| Orchestration | LangGraph v0.2 | Stateful agent graph — conditional routing, checkpointing, human-in-the-loop gates |
| Toolkit | LangChain v0.3 | WebBaseLoader, output parsers, LLM wrappers, text splitters |
| Primary LLM | Claude Sonnet 4.5 | Doc parsing, schema mapping, TypeScript code gen, test writing (4 of 5 agents) |
| Audit LLM | GPT-4o | Security audit only — different model prevents blind-spot overlap with code generator |
| RAG | Not required | API docs fetched live and fit in 200K context window. Retrieval adds latency with no benefit. |
| Backend | FastAPI + Python | LangGraph host, SSE streaming, credential vault integration |
| Observability | LangSmith | Per-agent trace logging, token usage, latency tracking |
3. Why LangGraph? Is It Necessary?
LangGraph manages a directed stateful graph of specialist agents. The graph branches conditionally — if the Security Auditor fails, it routes back to Code Generator rather than forward to Test Writer. LangChain provides utilities: document loaders, the unified ChatAnthropic/ChatOpenAI interface, and structured output parsers.
4. Why GPT-4o for Security Auditing Only?
This is adversarial diversity. A model reviewing code it just generated has the same blind spots as when it wrote it. GPT-4o brings a different training distribution, catching issues the generator missed. This is enforced in the LangGraph graph definition: the audit node cannot use the same model as the codegen node.
options to generate code