Large language models are increasingly expected to handle long-horizon tasks by orchestrating diverse tools. The new paper introduces ProgAgent, a method that learns tool orchestration through programmatic training, rather than relying on the more common Direct Tool Calling (DTC) approach.
In DTC, the model invokes tools sequentially, waiting for each response before deciding the next step. ProgAgent instead aims to generate a program that structures the orchestration, potentially improving efficiency and robustness over long horizons.
The abstract is truncated, so details on the training procedure and experimental results are not available from the source. The core contribution, however, is clear: a shift from direct, step-by-step tool calls to programmatic orchestration for LLM agents.