Long-horizon tasks expose a weakness in conventional tool-use loops: a single agent must decompose, execute, and verify everything within one trajectory. The LAGO framework, described in a new arXiv preprint, proposes an alternative by treating specialized language agents as reusable options that can be stored and drawn upon as needed.

Instead of forcing one agent to handle an entire task, LAGO builds a growing set of focused agents, each suited to a particular subtask. This shifts the burden from a monolithic loop to a library of components that can be combined flexibly.

The abstract is brief, so details on how options are grown or selected are not yet available. Still, the core idea—reusability over one-shot execution—offers a clear direction for improving agent performance on complex, multi-step problems.