Agentic GraphRAG

Part III. Action: From Firefighting to Foresight

Part III brings the agent to life. A knowledge graph and memory give the agent a brain, but a brain without a nervous system is inert. Part III builds the horizontal workflow graph and connects the agent to the outside world through tools, transforming it from a queryable knowledge base into a system that thinks, plans, and acts.

By the end of Part III, the DevOps agent has faced its first real challenge—a vague high-latency alert from an SRE—and worked through it end to end. It decomposes the problem into a multistep plan, traverses dependencies in the knowledge graph, queries live metrics through an orchestrated tool, and identifies the root cause. It has gone from being a passive map to a competent diagnostician.

• Chapter 5 covers graph-based reasoning (multi-hop traversal, PathRAG-style verified paths, adaptive retrieval) and DAG-based planning (decomposition, parallel execution, resilient failure management, dynamic replanning) and applies these concepts to constructing the agent’s first horizontal workflow graph.

• Chapter 6 covers MCP-based tool integration, graph-based tool representation and discovery, security and information flow control, and grammar-constrained structured output, enabling the agent to reliably interact with external systems and live data.