Samsung R&D Releases langgraph-goat for Explainable Multi-Agent AI Systems

Samsung R&D Institute India-Bangalore has released langgraph-goat, an open-source implementation of the Graph of Agentic Thoughts (GoAT) framework for LangGraph. The tool addresses the opacity of multi-agent AI systems by capturing the reasoning trace as decisions are made, rather than reconstructing them post-hoc.

How GoAT Works

GoAT captures the decision process in a directed, attributed graph that preserves relationships among planning, memory retrieval, delegation, and execution. When a request arrives, a coordinator generates multiple candidate plans, which a Plan Critic module scores from 0 to 10 based on user intent, context, and common sense.

Crucially, the system retains unchosen plans as counterfactuals or "ghost nodes" with explanations for their rejection. This allows the system to answer not just what happened, but why specific alternatives were discarded. The graph structure includes:

• Nodes: Agent-task executions enriched with reasoning traces.
• Edges: Delegation chains, causal dependencies, and memory transitions.
• Memory Nodes: Records of long-term memory retrieval and its relevance.
• Ghost Nodes: Rejected plans with their rejection justifications.

Performance and Evaluation

The framework was evaluated against four baselines, including raw log-trace explanation and multi-turn Chain-of-Thought (CoT) reasoning, using 2,000 dialogues from an in-house personal assistant and 1,000 from the MultiWOZ 2.4 benchmark. GoAT outperformed all baselines on every metric, with the largest gap observed in causality alignment, where it correctly captured cause-and-effect relationships between agents.

In a blind human evaluation involving seven independent researchers rating 50 test conversations, GoAT achieved a score of 4.71 out of 5 for impact on user trust, an improvement of 1.07 over the strongest baseline. This suggests that exposing considered alternatives and rejection reasons significantly enhances perceived credibility.

Implementation and Availability

The library integrates at the planner node, requiring no restructuring of existing LangGraph applications. It is written in Python, uses NetworkX for graph construction, and supports any LangChain-compatible model for the explanation layer. The source code is available on GitHub, and the package is published on PyPI.