DAG and Topological sort to reduce unnecessary reasoning by LLMs

This paper presents a novel orchestration architecture for multi-agent AI systems, specifically the OctaMind system. It replaces the traditional iterative ReAct loop with a "plan once, sort, execute deterministically" pattern. By invoking an LLM exactly once to construct a Directed Acyclic Graph (DAG) and using Kahn's topological sort for sequencing, the system reduces orchestration LLM calls by up to 70% on complex tasks. The architecture features a two-level design: a macro-DAG planner for routing tasks across heterogeneous agents and a micro-DAG engine for individual tool calls within sub-agents.

Performance Metrics:

1. Reduces LLM calls by 58–85% on multi-step workflows.
2. Topological sort time: <0.1 ms.
3. Planning success rate for single-agent tasks: ~98%.