LLM-as-Code: Agentic Programming for Agent Harness
The point
A paper proposes Agentic Programming, where the program governs control flow and the model is called only for reasoning or generation. The authors argue token explosion, control flow hallucination and unreliable completion follow from letting a probabilistic model handle looping, branching and sequencing.
Why it matters
For builders, control in code could make agent runs more reliable.
Key details
- LLM-as-Code keeps flexibility inside calls, but cannot change the execution path
- Context forms a directed acyclic graph from the call tree, sized by call depth
- Computer use study improved stability on visual sequences; arXiv 2606.15874
Original ~5 min → 25 sec here Source: arXiv cs.AI, Nov 9, 2026