# Graft AST Analysis Map: langchain-ai/open_deep_research
Repository: https://github.com/langchain-ai/open_deep_research
Tool: Graft v0.16.0 (Static AST Analysis)
Generated: 2026-09-28

## 1. Summary Statistics
- Total Files: 22
- Total Symbols: 180 (128 functions, 49 classes, 22 files, 3 methods)
- Total Edges: 485
- Tokens Saved: 89,013 tokens (73,264 from graft map, 15,749 from graft callers)

## 2. Directory Clusters & Hubs
- src/legacy/ (10 files, 95 symbols)
  Hubs:
  * _load_mcp_tools (multi_agent.py, 2 incoming edges)
  * azureaisearch_search_async (utils.py, 2 incoming edges)
  * get_research_tools (multi_agent.py, 2 incoming edges)
  * get_supervisor_tools (multi_agent.py, 2 incoming edges)
  * get_search_tool (multi_agent.py, 2 incoming edges)

- src/open_deep_research/ (5 files, 53 symbols)
  Hubs:
  * get_api_key_for_model (utils.py:L892-L914, 7 incoming edges)
  * is_token_limit_exceeded (utils.py:L665-L701, 3 incoming edges)
  * _find_mcp_error_in_exception_chain (utils.py:L399-L409, 2 incoming edges)
  * get_all_tools (utils.py:L569-L597, 2 incoming edges)
  * deep_researcher.py (researcher, researcher_tools)

- tests/ (6 files, 25 symbols)
  Hubs:
  * _format_input_query (evaluators.py:L12-L22, 5 incoming edges)
  * extract_langsmith_data (extract_langsmith_data.py, 1 incoming edge)
  * main (extract_langsmith_data.py, 1 incoming edge)
  * supervisor_parallel_evaluation.py
  * pairwise_evaluation.py

- src/security/ (1 file, 7 symbols)
  Hubs:
  * verify_token (auth.py, 1 incoming edge)

## 3. Key Architectural Insights for Hermes Agent Stack
1. Recursive Multi-Agent Research Loops: DeepResearcher coordinates supervisor and worker agents with parallel tool invocation.
2. Dynamic MCP Tool Integration: Seamless MCP loading and error recovery in exception chains (_find_mcp_error_in_exception_chain).
3. Token Budget Guardrails: is_token_limit_exceeded guards context window overflow before triggering report synthesis.