Welcome to the new NetworkX-First implementation. This version replaces the manual graph management with a robust DAG structure and integrates modern tools.
- File:
core/loop.py - Logic: Uses
networkx.DiGraphto manage the plan. - Context: Strictly isolated. Node B only sees Node A's output if
A -> B. - Summarizer: Special agent with Global Read Access to synthsize everything.
- Browser: Hybrid Mode.
- Fast Search:
duckduckgo(Text). - Deep Action:
browser-use(Vision).
- Fast Search:
- Memory: Dual Layer.
- Short-term: Session Context (NetworkX attributes).
- Long-term:
mem0(User Profile, Local).
- Sandbox:
e2bcapable (currently wrapped intools/sandbox.py).
16_NetworkX/
├── agents/ # AgentRunners (Logic)
├── config/ # YAML Configs + Models
├── core/ # Main Loop & Utils
├── memory/ # Context & Mem0 Store
├── mcp_servers/ # Tool Servers (Browser, RAG)
├── prompts/ # System Prompts (Markdown)
├── tools/ # Helper Scripts
└── app.py # Entry Point
uv run 16_NetworkX/app.pyType your query (e.g., "Plan a 3 day trip to Tokyo").
uv run 16_NetworkX/test_run.py- Agents:
config/agent_config.yamlmaps Agents -> Prompts -> Tools. - Models:
config/models.jsondefines LLMs (Gemini, Ollama). - Prompts: Edit
prompts/*.mdto change agent behavior.
- UI: Connect
ui/visualizer.pyto a React Flow frontend. - E2B: Fully replace
tools/sandbox.pywith E2B SDK. - Mem0: Enable active learning in
SummarizerAgent.