Reducing AI Agent Token Usage with Graft Knowledge Graphs

The Breakthrough

Graft optimizes AI coding agents by replacing standard terminal-based file searching with a persistent knowledge graph, which maps code nodes and dependencies to provide the agent with precise file locations before it executes tool calls.

What Actually Worked

• Knowledge Graph Mapping: The tool parses the project to create a JSON-based map of code components, where nodes represent files or functions and edges represent functional dependencies, allowing the agent to understand relationships that vector search misses.
• Automated Hook Integration: Running graft init installs scripts that automatically inject context into the agent session, including a hook that attaches up to three relevant file locations to every prompt and a post-edit hook that updates the map in real-time.
• Hybrid Execution Modes: Users can leverage both a CLI-based hook system for faster performance and an MCP (Model Context Protocol) server that allows the agent to query the map only when necessary, balancing speed with accuracy.
• Context-Aware Initialization: By running graft build on existing projects, the tool maps the codebase to ensure the agent has immediate access to structural data without requiring manual indexing or external API keys.

Before / After

• Task Completion Time: 47 minutes (without Graft) vs. 39 minutes (with Graft).
• Context Window Usage: 35% (without Graft) vs. 31% (with Graft).
• Benchmark Averages (162 runs): 60% less time, 46% fewer tool calls, 42% fewer tokens, and 32% lower cost.

Context

Coding agents typically burn through tokens and context windows by repeatedly searching the file system using terminal commands to locate relevant code. Because these agents often fail to find the correct file on the first attempt, they trigger multiple turns of tool usage, each of which resends the entire conversation history. Graft mitigates this by providing a structural map of the project, ensuring the agent retrieves only the necessary code segments rather than relying on iterative, trial-and-error searches.

Content References

{"type": "tool", "title": "Graft", "url": "https://trailhq.com/graft", "context": "recommended"}, {"type": "tool", "title": "Hedra", "url": "https://hedra.link/ailabs", "context": "mentioned"}