Automatic discovery of optimal task decompositions

Investigate methods to automatically discover optimal decompositions of long-horizon tasks into step-level subtasks suitable for execution by large language model agents, rather than relying on a priori human-defined step boundaries.

Background

The paper’s Massively Decomposed Agentic Processes (MDAPs) framework assumes a task can be partitioned into minimal steps that LLM microagents can execute reliably, enabling per-step error correction. While the work focuses on execution under a fixed step definition, the authors note that the granularity of decomposition critically affects scalability and reliability.

They explicitly point out that determining how to discover such decompositions automatically is unresolved, highlighting the need for algorithmic approaches that identify optimal step boundaries that maintain high per-step success rates and enable effective voting-based error correction.