DKGI paper: agents write their own tools to search knowledge graphs, beating baselines by 6.7 Hit@1
yuz9yuz · x · 2026-10-16
The arXiv paper 'Knowledge Graphs as Files' proposes DKGI (direct knowledge graph interaction): instead of predefined graph APIs, agents receive full node/edge files and write Python plus shell code to follow relations and select evidence, creating and revising tools on the fly. Execution feedback guides tool revision, large results stay in files rather than the context window, and training-free DKGI beats the strongest published baseline by at least 6.7 Hit@1 on every benchmark graph. Ablations show multi-turn programming beats shell-only access and predefined tools. The reposter notes his prior startup found bash-based agentic search more efficient than MCP-style graph tools.
More from coding & agent
- What are cloud agent swarms actually for? Real use cases from federal courts to overnight testing — Dan_Jeffries1 · 2026-10-16
- Codex reads across chat windows to self-test whether a skill fired — lxfater · 2026-10-16
- Agent-written code team cut pages from 14/month to 2 with AI review gate — Few-Pilot7575 · 2026-10-16
- DHH's Omarchy Shenzhen Meetup Draws 400+ With Demos on Xiaomi, ARM and iPhone Ports — LoganGrasby · 2026-10-16
- Multi-Hop RAG Failures Come From Search Drift, Not Retrieval Limits — _reachsumit · 2026-10-16
- ShadHub Launches as a Directory for Browsing and Discovering GitHub Registries — shadcn · 2026-10-16