LLM-as-Code Agentic Programming for Agent Harness

2026-06-14

AI Summary

Traditional LLM agents delegate control flow to probabilistic language models, often leading to issues such as token explosion, control-flow hallucination, and unreliable execution. This work proposes the Agentic Programming paradigm, wherein program logic fully governs control flow and invokes the language model strictly as a reasoning or generation subroutine—termed “LLM-as-Code”—only when necessary. This design ensures stable and predictable execution paths through program-driven control, depth-aware context management, and execution history modeled as a directed acyclic graph (DAG). Evaluated on computer operation agent tasks, the approach demonstrates significantly enhanced stability and reliability in long-sequence visual manipulation, thereby validating the efficacy of the proposed architecture.

Abstract

Every major LLM agent framework gives the LLM the role of orchestrator; the model decides what to do next, when to call tools, and when to stop. We argue that token explosion, control-flow hallucination, and unreliable completion are not implementation bugs but architectural consequences of assigning the deterministic work of looping, branching, and sequencing to a probabilistic system. A better prompt or a stronger model cannot guarantee the reliability of the LLM agent. We therefore propose Agentic Programming, in which the program governs all control flow, and the LLM is itself part of it, an adaptive component we call LLM-as-Code and invoke only where a task calls for reasoning or generation. Within each call the model keeps full flexibility, but it cannot alter the program's execution path. With control in the program, the LLM's context is built from the execution history's call tree and forms a directed acyclic graph (DAG). Each call's context length is then determined by its call depth rather than by accumulation over steps. A case study of computer-use agents shows that the design is practical, not just a theoretical stance, substantially improving the stability of long visual operation sequences.

Problem

• LLM agent
• control-flow hallucination
• token explosion
• reliable execution
• agentic programming

Innovation

• Agentic Programming
• LLM-as-Code
• control flow
• directed acyclic graph
• reliable agent execution