Coding Agents with Environment Interaction: A Theoretical Perspective
Nicolas Menet, Michael Hersche, Andreas Krause, Abbas Rahimi
TL;DR
This work provides a probabilistic framework for coding agents that interact with execution environments, addressing two main paradigms: post-generation selection and in-generation backprompting. It shows that using functional similarity to group behavior yields a higher signal-to-noise ratio than strict functional equivalence, thereby offering a stronger inductive bias for selecting correct code. It also treats backprompting as an in-context approximation to Thompson sampling and derives a regret bound with an irreducible component due to task-description ambiguity, explaining why environment feedback cannot completely overcome misalignment. Across three open-weight models and multiple datasets, the authors validate that soft (similarity-based) estimators consistently outperform hard (equivalence-based) ones, and that backprompting is most effective when the unobservable reward component is small or the task description is clarified; they further introduce QiskitHumanEvalSimX to probe improvements in task descriptions. These insights guide practical design choices for task descriptions and feedback processing, highlighting the trade-offs between computation, context length, and the quality of chosen evaluation signals in real-world software engineering with LLMs.
Abstract
Coding agents are increasingly utilized in test-driven software development, yet the theoretical mechanisms behind their environment-interaction strategies remain underexplored. We provide a probabilistic framework for two dominant paradigms: code selection after generation using the execution environment, and code generation conditioned on environment feedback. First, we formalize several well-established selection heuristics as environment-aware estimators of code correctness. We theoretically prove that estimators based on fuzzy functional similarity add an inductive bias and strictly dominate estimators based on functional equivalence in terms of signal-to-noise ratio. Second, we frame backprompting as an in-context approximation of Thompson sampling. We derive a novel regret bound for reward functions with unobservable components, theoretically explaining why the effectiveness of backprompting is limited by the ambiguity of the informal task description (an irreducible regret). Using three state-of-the-art open weight models, we corroborate these findings across BigCodeBenchHard, LeetCodeDataset, and QiskitHumanEvalSim. Our formalization also suggests how to improve task descriptions effectively, leading to a new benchmark, QiskitHumanEvalSimX.