LLM-as-Code Agentic Programming for Agent Harness

The paper introduces Agentic Programming, a paradigm where the control flow of LLM agents is governed by a program rather than the LLM itself, addressing issues like token explosion and control-flow hallucination. The proposed LLM-as-Code framework allows the LLM to function as an adaptive component, invoked only for reasoning or generation tasks, while maintaining the program's execution path. This approach, demonstrated through a case study on computer-use agents, significantly enhances the stability of complex operation sequences by structuring context as a directed acyclic graph (DAG) based on call depth.