Deep Research Agent - Enhanced with OpenSWE Plan

[rahulramakrishnanprema]

Deep Research Agent - Enhanced with OpenSWE Plan

1. Project Setup

   • Install LangChain, LangGraph, OpenSWE, HuggingFace Transformers
   • Set up OpenSWE's multi-agent architecture
   • Configure MCP protocol endpoints

2. User Input Handling (Manager Agent)

   • OpenSWE Manager receives and routes user queries
   • Initializes research session state

3. Research Planning (Planner Agent)

   • Uses HuggingFace LLM to analyze user query
   • Creates detailed, step-by-step research execution plan
   • Generates optimized search queries

4. Live Web Search Execution (Researcher Agent)

   • Executes search queries against web APIs
   • Operates in secure sandboxed environment
   • Collects real-time search results

5. Data Evaluation Loop (LangGraph Orchestration)

   • Uses OpenSWE's workflow nodes architecture
   • Evaluates research completeness
   • Loops until sufficient data gathered

6. Summary Generation (Reviewer Agent)

   • Validates collected research data
   • Synthesizes comprehensive summary with citations
   • Reviews for quality and completeness

7. Communication Setup (MCP Protocol)

   • Uses OpenSWE's built-in MCP integration
   • Enables communication between all agents
   • Manages state sharing across workflow

8. Output Results

   • Returns final research summary
   • Maintains conversation context for follow-up

Agent Context

{ "tasks": [ { "id": "cb3f7909-2135-4278-b6a4-53eb973761d4", "taskIndex": 0, "request": "[original issue]\n**Deep Research Agent - Enhanced with OpenSWE Plan**\nDeep Research Agent - Enhanced with OpenSWE Plan\n\n1. Project Setup\n - Install LangChain, LangGraph, OpenSWE, HuggingFace Transformers\n - Set up OpenSWE's multi-agent architecture\n - Configure MCP protocol endpoints\n\n2. User Input Handling (Manager Agent)\n - OpenSWE Manager receives and routes user queries\n - Initializes research session state\n\n3. Research Planning (Planner Agent)\n - Uses HuggingFace LLM to analyze user query\n - Creates detailed, step-by-step research execution plan\n - Generates optimized search queries\n\n4. Live Web Search Execution (Researcher Agent)\n - Executes search queries against web APIs\n - Operates in secure sandboxed environment\n - Collects real-time search results\n\n5. Data Evaluation Loop (LangGraph Orchestration)\n - Uses OpenSWE's workflow nodes architecture\n - Evaluates research completeness\n - Loops until sufficient data gathered\n\n6. Summary Generation (Reviewer Agent)\n - Validates collected research data\n - Synthesizes comprehensive summary with citations\n - Reviews for quality and completeness\n\n7. Communication Setup (MCP Protocol)\n - Uses OpenSWE's built-in MCP integration\n - Enables communication between all agents\n - Manages state sharing across workflow\n\n8. Output Results\n - Returns final research summary\n - Maintains conversation context for follow-up", "title": "Deep Research Agent Enhanced with OpenSWE Plan", "createdAt": 1756043922353, "completed": false, "planRevisions": [ { "revisionIndex": 0, "plans": [ { "index": 0, "plan": "Update `requirements.txt` to include `OpenSWE` and `transformers` libraries. Ensure all existing dependencies (`langchain`, `langgraph`) are also present.", "completed": false }, { "index": 1, "plan": "Create a new Python file, `deep_research_agent.py`, to house the LangGraph multi-agent architecture. This file will define the state graph, nodes for each agent (Manager, Planner, Researcher, Reviewer), and the conditional edges for orchestration. Refer to LangGraph's multi-agent system documentation (e.g., `https://langchain-ai.github.io/langgraph/tutorials/multi_agent/agent_supervisor/`) for guidance on structuring the supervisor and worker agents.", "completed": false }, { "index": 2, "plan": "Implement the Manager Agent within `deep_research_agent.py`. This agent will be the entry point for user queries, initialize the research session state, and route the query to the Planner Agent. The session state should be a dictionary or Pydantic model to store conversation history, research plan, and collected data.", "completed": false }, { "index": 3, "plan": "Implement the Planner Agent within `deep_research_agent.py`. This agent will utilize a HuggingFace LLM (e.g., by integrating `transformers` library) to analyze the user query, create a detailed step-by-step research execution plan, and generate optimized search queries. The plan and search queries should be stored in the session state.", "completed": false }, { "index": 4, "plan": "Implement the Researcher Agent within `deep_research_agent.py`. This agent will execute the generated search queries against web APIs (e.g., using a tool like Tavily Search Engine, as referenced in `https://langchain-ai.github.io/langgraph/tutorials/get-started/2-add-tools/`). It should operate in a secure sandboxed environment (if external API calls are made) and collect real-time search results, storing them in the session state.", "completed": false }, { "index": 5, "plan": "Implement the Data Evaluation Loop using LangGraph's workflow nodes architecture within `deep_research_agent.py`. This loop will evaluate the completeness of the collected research data and continue to loop, potentially re-invoking the Planner and Researcher agents, until sufficient data is gathered. Refer to LangGraph's core concepts on States, Nodes, and Edges (e.g., `https://langchain-ai.github.io/langgraph/concepts/low_level/`) for structuring this loop.", "completed": false }, { "index": 6, "plan": "Implement the Reviewer Agent within `deep_research_agent.py`. This agent will validate the collected research data, synthesize a comprehensive summary with citations, and review it for quality and completeness. The final summary should be stored in the session state.", "completed": false }, { "index": 7, "plan": "Configure the MCP protocol endpoints for communication between all agents within `deep_research_agent.py`. Utilize the `langchain-mcp-adapters` library as described in the LangGraph documentation (e.g., `https://langchain-ai.github.io/langgraph/agents/mcp/` and `https://langchain-ai.github.io/langgraph/concepts/server-mcp/`). This will enable state sharing and seamless interaction across the workflow.", "completed": false }, { "index": 8, "plan": "Modify `main.py` to import and run the `deep_research_agent.py` as the primary application. The `main.py` file should handle initial user input and display the final research summary returned by the Deep Research Agent.", "completed": false } ], "createdAt": 1756043922353, "createdBy": "agent" } ], "activeRevisionIndex": 0 } ], "activeTaskIndex": 0 } [ "Update `requirements.txt` to include `OpenSWE` and `transformers` libraries. Ensure all existing dependencies (`langchain`, `langgraph`) are also present.", "Create a new Python file, `deep_research_agent.py`, to house the LangGraph multi-agent architecture. This file will define the state graph, nodes for each agent (Manager, Planner, Researcher, Reviewer), and the conditional edges for orchestration. Refer to LangGraph's multi-agent system documentation (e.g., `https://langchain-ai.github.io/langgraph/tutorials/multi_agent/agent_supervisor/`) for guidance on structuring the supervisor and worker agents.", "Implement the Manager Agent within `deep_research_agent.py`. This agent will be the entry point for user queries, initialize the research session state, and route the query to the Planner Agent. The session state should be a dictionary or Pydantic model to store conversation history, research plan, and collected data.", "Implement the Planner Agent within `deep_research_agent.py`. This agent will utilize a HuggingFace LLM (e.g., by integrating `transformers` library) to analyze the user query, create a detailed step-by-step research execution plan, and generate optimized search queries. The plan and search queries should be stored in the session state.", "Implement the Researcher Agent within `deep_research_agent.py`. This agent will execute the generated search queries against web APIs (e.g., using a tool like Tavily Search Engine, as referenced in `https://langchain-ai.github.io/langgraph/tutorials/get-started/2-add-tools/`). It should operate in a secure sandboxed environment (if external API calls are made) and collect real-time search results, storing them in the session state.", "Implement the Data Evaluation Loop using LangGraph's workflow nodes architecture within `deep_research_agent.py`. This loop will evaluate the completeness of the collected research data and continue to loop, potentially re-invoking the Planner and Researcher agents, until sufficient data is gathered. Refer to LangGraph's core concepts on States, Nodes, and Edges (e.g., `https://langchain-ai.github.io/langgraph/concepts/low_level/`) for structuring this loop.", "Implement the Reviewer Agent within `deep_research_agent.py`. This agent will validate the collected research data, synthesize a comprehensive summary with citations, and review it for quality and completeness. The final summary should be stored in the session state.", "Configure the MCP protocol endpoints for communication between all agents within `deep_research_agent.py`. Utilize the `langchain-mcp-adapters` library as described in the LangGraph documentation (e.g., `https://langchain-ai.github.io/langgraph/agents/mcp/` and `https://langchain-ai.github.io/langgraph/concepts/server-mcp/`). This will enable state sharing and seamless interaction across the workflow.", "Modify `main.py` to import and run the `deep_research_agent.py` as the primary application. The `main.py` file should handle initial user input and display the final research summary returned by the Deep Research Agent." ]

[open-swe]

🟠 Plan Ready for Approval 🟠

I've generated a plan for this issue and it's ready for your review.

Plan: Deep Research Agent Enhanced with OpenSWE Plan

• Task 1:

Update `requirements.txt` to include `OpenSWE` and `transformers` libraries. Ensure all existing dependencies (`langchain`, `langgraph`) are also present.

• Task 2:

Create a new Python file, `deep_research_agent.py`, to house the LangGraph multi-agent architecture. This file will define the state graph, nodes for each agent (Manager, Planner, Researcher, Reviewer), and the conditional edges for orchestration. Refer to LangGraph's multi-agent system documentation (e.g., `https://langchain-ai.github.io/langgraph/tutorials/multi_agent/agent_supervisor/`) for guidance on structuring the supervisor and worker agents.

• Task 3:

Implement the Manager Agent within `deep_research_agent.py`. This agent will be the entry point for user queries, initialize the research session state, and route the query to the Planner Agent. The session state should be a dictionary or Pydantic model to store conversation history, research plan, and collected data.

• Task 4:

Implement the Planner Agent within `deep_research_agent.py`. This agent will utilize a HuggingFace LLM (e.g., by integrating `transformers` library) to analyze the user query, create a detailed step-by-step research execution plan, and generate optimized search queries. The plan and search queries should be stored in the session state.

• Task 5:

Implement the Researcher Agent within `deep_research_agent.py`. This agent will execute the generated search queries against web APIs (e.g., using a tool like Tavily Search Engine, as referenced in `https://langchain-ai.github.io/langgraph/tutorials/get-started/2-add-tools/`). It should operate in a secure sandboxed environment (if external API calls are made) and collect real-time search results, storing them in the session state.

• Task 6:

Implement the Data Evaluation Loop using LangGraph's workflow nodes architecture within `deep_research_agent.py`. This loop will evaluate the completeness of the collected research data and continue to loop, potentially re-invoking the Planner and Researcher agents, until sufficient data is gathered. Refer to LangGraph's core concepts on States, Nodes, and Edges (e.g., `https://langchain-ai.github.io/langgraph/concepts/low_level/`) for structuring this loop.

• Task 7:

Implement the Reviewer Agent within `deep_research_agent.py`. This agent will validate the collected research data, synthesize a comprehensive summary with citations, and review it for quality and completeness. The final summary should be stored in the session state.

• Task 8:

Configure the MCP protocol endpoints for communication between all agents within `deep_research_agent.py`. Utilize the `langchain-mcp-adapters` library as described in the LangGraph documentation (e.g., `https://langchain-ai.github.io/langgraph/agents/mcp/` and `https://langchain-ai.github.io/langgraph/concepts/server-mcp/`). This will enable state sharing and seamless interaction across the workflow.

• Task 9:

Modify `main.py` to import and run the `deep_research_agent.py` as the primary application. The `main.py` file should handle initial user input and display the final research summary returned by the Deep Research Agent.

Please review the plan and let me know if you'd like me to proceed, make changes, or if you have any feedback.

✅ Plan Accepted ✅

The proposed plan was accepted.

Plan: Deep Research Agent Enhanced with OpenSWE Plan

• Task 1:

Update `requirements.txt` to include `OpenSWE` and `transformers` libraries. Ensure all existing dependencies (`langchain`, `langgraph`) are also present.

• Task 2:

Create a new Python file, `deep_research_agent.py`, to house the LangGraph multi-agent architecture. This file will define the state graph, nodes for each agent (Manager, Planner, Researcher, Reviewer), and the conditional edges for orchestration. Refer to LangGraph's multi-agent system documentation (e.g., `https://langchain-ai.github.io/langgraph/tutorials/multi_agent/agent_supervisor/`) for guidance on structuring the supervisor and worker agents.

• Task 3:

Implement the Manager Agent within `deep_research_agent.py`. This agent will be the entry point for user queries, initialize the research session state, and route the query to the Planner Agent. The session state should be a dictionary or Pydantic model to store conversation history, research plan, and collected data.

• Task 4:

Implement the Planner Agent within `deep_research_agent.py`. This agent will utilize a HuggingFace LLM (e.g., by integrating `transformers` library) to analyze the user query, create a detailed step-by-step research execution plan, and generate optimized search queries. The plan and search queries should be stored in the session state.

• Task 5:

Implement the Researcher Agent within `deep_research_agent.py`. This agent will execute the generated search queries against web APIs (e.g., using a tool like Tavily Search Engine, as referenced in `https://langchain-ai.github.io/langgraph/tutorials/get-started/2-add-tools/`). It should operate in a secure sandboxed environment (if external API calls are made) and collect real-time search results, storing them in the session state.

• Task 6:

Implement the Data Evaluation Loop using LangGraph's workflow nodes architecture within `deep_research_agent.py`. This loop will evaluate the completeness of the collected research data and continue to loop, potentially re-invoking the Planner and Researcher agents, until sufficient data is gathered. Refer to LangGraph's core concepts on States, Nodes, and Edges (e.g., `https://langchain-ai.github.io/langgraph/concepts/low_level/`) for structuring this loop.

• Task 7:

Implement the Reviewer Agent within `deep_research_agent.py`. This agent will validate the collected research data, synthesize a comprehensive summary with citations, and review it for quality and completeness. The final summary should be stored in the session state.

• Task 8:

Configure the MCP protocol endpoints for communication between all agents within `deep_research_agent.py`. Utilize the `langchain-mcp-adapters` library as described in the LangGraph documentation (e.g., `https://langchain-ai.github.io/langgraph/agents/mcp/` and `https://langchain-ai.github.io/langgraph/concepts/server-mcp/`). This will enable state sharing and seamless interaction across the workflow.

• Task 9:

Modify `main.py` to import and run the `deep_research_agent.py` as the primary application. The `main.py` file should handle initial user input and display the final research summary returned by the Deep Research Agent.

Proceeding to implementation...