Intel / article

ECLoop: Evidence-Conditioned Execution Layer for Coding Agents

Source: arxiv.org
Author: Yisen Xu, Chenglin Li, Zehao Wang, Jinqiu Yang, Tse-Hsun Chen
Date: Published Aug 3, 2026

Context

ECLoop places a control layer between a coding and its repository, delaying edits or submission until required observations have been made. It derives those requirements from the issue description and repository structure, then tracks whether the agent's actions satisfy them. Across 500 tasks, two models, and two agent frameworks, the authors report first-attempt success gains of 4.8 to 11.8 percentage points without retraining. The proposal makes investigation a condition of execution. It complements checking whether passing tests distinguish the original bug, which examines the evidence supporting a repair.

Key takeaways · AI-distilled

ECLoop derives its conditions per task from the issue text plus repo structure, tracks which ones the agent's trajectory has actually satisfied, and postpones any action whose preconditions are still unmet. It wraps the model instead of changing it.
Structure beat prose in the ablation: giving the agent an equivalent natural-language summary of the same guidance worked worse than the machine-checkable condition list. The enforcement is doing the work, not the advice.
The result held across two different models and two different agent scaffolds over all 500 SWE-bench Verified instances, which argues the gain comes from the execution layer rather than one lucky model-and- pairing.
Blocking an action early is cheaper than letting it run: cutting off an unsupported edit also cuts the whole dead-end branch of work that would have followed it.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.

Key quotes

“LLM-based coding agents often edit source code or submit patches before examining enough repository evidence to justify the change, a failure pattern we call premature commitment.”

“Evaluated on all 500 instances of SWE-bench Verified with two language models and two agent scaffolds, ECLoop raises Pass@1 by 4.8-11.8 percentage points without model retraining or scaffold changes.”

“These gains come at no additional inference cost: by redirecting the agent before it pursues unsupported actions, ECLoop lowers average token consumption by up to 12.1%.”

Read the source arxiv.org