Agent Experience Graph: An Execution Record and Analysis System for Reusable AI Agent Workflows

Agent Experience Graph is an open-source tool designed to address issues in AI agent development and operation, such as insufficient observability, difficulty in experience accumulation, and debugging challenges. By modeling the agent execution process as a knowledge graph, it enables the recording, visualization, analysis, and reuse of execution trajectories, supporting continuous workflow optimization and the accumulation of agent capabilities.

With the improvement of LLM capabilities, AI agents are moving towards production deployment, but they face three major challenges:

1. Insufficient Observability: Agent behaviors are highly dynamic, and traditional logs struggle to fully capture complex execution processes;
2. Difficulty in Experience Accumulation: Valuable experiences from successful executions (e.g., tool combinations, error correction strategies) are lost after a single execution;
3. Debugging Difficulties: It is hard to trace root causes (e.g., wrong tool selection, context deviation) when execution fails. The Agent Experience Graph project was born to solve these problems by modeling the execution process as a graph.

The core of the project is modeling agent execution as an experience graph:

• Data Model: Includes execution sessions (meta-information of complete task execution), execution steps (atomic operations like LLM reasoning, tool calls), decision nodes (records of multi-path choices), and state snapshots (complete state at key nodes);
• Graph Construction: Uses a property graph model with node types such as Session, Step, Tool, and edge types like HAS_STEP, CALLS, etc., to structure the execution process for flexible query and analysis.

The system consists of three core modules:

1. Execution Recorder: A lightweight SDK that automatically embeds points to capture key events, supporting context propagation and real-time/batch reporting;
2. Analysis Engine: Mines execution patterns, detects anomalies, clusters similar executions, and analyzes bottlenecks;
3. Knowledge Reuse Layer: Performs case retrieval, strategy recommendation, and generates workflow templates, transforming experiences into reusable assets.

Technical implementation details:

• Storage Layer: A layered storage strategy, with graph databases (Neo4j) storing topological structures, time-series databases storing metrics, and object storage storing large-volume data;
• Query Interface: Supports Cypher graph queries, natural language-to-query (lowering the threshold), and visual interfaces to display execution trajectories, etc.

Application scenarios include:

1. Development and Debugging: View execution trajectories to quickly locate problems;
2. Continuous Optimization: Establish a closed loop of collection-analysis-improvement-verification;
3. Team Collaboration: Unify views to accumulate best practices;
4. Compliance Audit: Provide complete audit logs and identify security risks.

Agent Experience Graph addresses the pain points of agent observability and experience management, laying the foundation for continuous learning and capability evolution. Future directions: Support multi-agent collaboration analysis, introduce reinforcement learning to extract optimal strategies, and integrate with AutoML to automatically optimize architectures.