Current document formats, designed for linear human consumption, are fundamentally misaligned with how autonomous LLM agents operate. Agents retrieve information, not read linearly. This mismatch forces inefficient token usage, state compounding, and indiscriminate information broadcasting within multi-agent systems. The authors propose that this is not a limitation of prompt engineering, retrieval, or compression, but an intrinsic format problem.
Reimagining Documents as Knowledge Graphs
Introducing the OBJECTGRAPH file format (.og), a radical departure from text-based documents. OBJECTGRAPH structures information as a typed, directed knowledge graph, optimized for traversal by AI agents. Crucially, it is a strict superset of Markdown, ensuring human readability and compatibility with existing .md files without requiring new infrastructure beyond a simple query protocol. This innovation directly tackles the Document Consumption Problem by satisfying six critical structural properties that existing formats fail to meet simultaneously.
Native Primitives for Intelligent Agent Interaction
The OBJECTGRAPH file format introduces novel, built-in primitives designed for advanced agent capabilities. These include the Progressive Disclosure Model, the Role-Scoped Access Protocol, and Executable Assertion Nodes. These native elements enable more nuanced and efficient agent interactions. Empirical evaluations across diverse document classes and agent task types demonstrate remarkable efficiency gains, with token reductions up to 95.3% observed, and task accuracy remaining statistically indistinguishable from baseline performance (p > 0.05). Furthermore, a transpiler achieves 98.7% content preservation on a held-out benchmark, highlighting the robustness of the format conversion.