Glassmind Navigator

💡 Inspiration Modern AI agents make decisions that even their creators can’t fully see. They act, remember, reason, take actions—and yet their internal steps are a black box. Track B of the Great Agent Hackathon asks a simple question: 👉 “Can you build an agent you can actually understand?” That challenge matches perfectly with my long-term work on AHI (Artificial Human Intelligence), cognitive models, and agent architecture.

I wanted to build something that: Shows the entire reasoning flow Makes memory fully trackable Reveals hidden behavior patterns Helps developers debug, trust, and improve agents That became GlassMind Navigator.

🧠 What It Does GlassMind Navigator turns any AI agent into a transparent, explainable system. 🔍 Key Features

1. Full Trajectory Visualization Every step is captured: reasoning → tool calls → outputs → next state.
2. Memory Timeline Analyzer Shows memory_before and memory_after for every node Reveal how memory evolves across planning, research, and answering.
3. Behavior Pattern Analyzer Automatically discovers: tool overuse loops node frequency unnecessary steps hallucination triggers
4. Error Replay Mode Finds failure points and reconstructs what happened at the exact step that went wrong.
5. Before/After Optimization You change a config → run again → see how behavior improves. Perfect for showing reliability improvements.
6. Real-Time LangSmith Observability All nodes are traced using @traceable, making every internal operation auditable. 🛠 How We Built It 🧩 Architecture User ↓ LangGraph-style Agent (Plan → Research → Answer) ↓ Trajectory Recorder ↓ Streamlit Trace Visualizer ↓ Memory Timeline Analyzer ↓ Behavior Pattern Analyzer ↓ Improvement Engine

⚙️ Tech Used LangSmith traceable decorators for full observability Gemini / Valyu API for reasoning + web search Custom structured StepLog for memory + tool tracking Streamlit for interactive visualization Pandas for behavior pattern analytics JSON traces for reproducibility

🗂 Key Components graph.py → agent logic + instrumented nodes traces_loader.py → save/load full trajectories app.py → Streamlit UI behavior_analysis.py → statistical insights memory_analysis.py → future memory diffing 🚧 Challenges We Ran Into LLM timeouts from Ollama → solved by switching to Gemini Keeping LangSmith traces serializable Import path issues in Streamlit Handling memory_before / memory_after cleanly Ensuring state mutation doesn’t break trace consistency Designing explanations that are simple yet expressive But each challenge actually became part of the final story: making agents robust and explainable under chaos.

🏆 Accomplishments We’re Proud Of Built full observability without LangGraph’s built-in tools Generated actionable behavior insights from real traces Achieved end-to-end transparency: reasoning → tools → memory → patterns → failures Created a demo where judges can see: Why the agent made a decision What it remembered What it forgot What step caused an error

How behavior changed after improvements

Successfully integrated LangSmith + custom visualizations

Built a clean foundational framework you can extend to ANY agent

📚 What We Learned

Observability isn’t a plugin—it's a design philosophy Memory tracking is essential for agent debugging Tools shape behavior far more than we expect

Small config changes can drastically alter agent trajectories

Explainability is NOT automatic—it's something you engineer

Most importantly:

👉 Transparency is the missing link between agents and trust.

🚀 What’s Next for GlassMind Navigator Coming Improvements:

Visual DAG-style graph of agent trajectories

Step diffs with natural-language explanations

Multi-agent support

Automated anomaly detection

Config optimizers using RL

Support for Bedrock Agents & AWS Strands

Plugin system for custom tools and memory managers

Long-term Vision:

A universal observability dashboard for all agents— bringing Trust, Safety, and Transparency to real production agentic systems.

Built With

• google-gemini
• langchain
• langsmith
• python
• streamlit