SAFEdit: Does Multi-Agent Decomposition Resolve the Reliability Challenges of Instructed Code Editing?

2026-04-28

AI Summary

Large language models often fail in instruction-driven code editing tasks due to hallucination or neglect of test constraints. To address this, this work proposes SAFEdit, a multi-agent framework that decomposes the editing process into three specialized roles—planning, editing, and verification—and introduces a Failure Abstraction Layer (FAL) to enable iterative refinement grounded in actual test feedback. By integrating visibility-aware planning, a minimal literal modification strategy, and multi-agent coordination, SAFEdit significantly enhances editing reliability and supports fine-grained failure attribution. Evaluated on EditBench, SAFEdit achieves a task success rate of 68.6%, outperforming single-model baselines by 3.8 percentage points and ReAct-based single-agent approaches by 8.6 points, with its iterative mechanism alone contributing a 17.4-point performance gain.

Abstract

Instructed code editing is a significant challenge for large language models (LLMs). On the EditBench benchmark, 39 of 40 evaluated models obtain a task success rate (TSR) below 60 percent, highlighting a gap between general code generation and the ability to perform instruction-driven editing under executable test constraints. To address this, we propose SAFEdit, a multi-agent framework for instructed code editing that decomposes the editing process into specialized roles to improve reliability and reduce unintended code changes. A Planner Agent produces an explicit, visibility-aware edit plan, an Editor Agent applies minimal, literal code modifications, and a Verifier Agent executes real test runs. When tests fail, SAFEdit uses a Failure Abstraction Layer (FAL) to transform raw test logs into structured diagnostic feedback, which is fed back to the Editor to support iterative refinement. We compare SAFEdit against both prior single-model results reported for EditBench and an implemented ReAct single-agent baseline under the same evaluation conditions. We used EditBench to evaluate SAFEdit on 445 code editing instances in five languages (English, Polish, Spanish, Chinese, and Russian) under varying spatial context variants. SAFEdit achieved 68.6 percent TSR, outperforming the single-model baseline by 3.8 percentage points and the ReAct single-agent baseline by 8.6 percentage points. The iterative refinement loop was found to contribute 17.4 percentage points to SAFEdit's overall success rate. SAFEdit's automated error analysis further indicates a reduction in instruction-level hallucinations compared to single-agent approaches, providing an additional framework component for interpreting failures beyond pass or fail outcomes.

Problem

• instructed code editing
• reliability
• large language models
• executable test constraints
• task success rate

Innovation

• multi-agent decomposition
• code editing
• Failure Abstraction Layer
• iterative refinement
• instruction-driven editing