Opera: A Verbal Critic Framework for Long-horizon Coding Agents

2026-09-27

AI Summary

This study addresses the limitations of existing long-horizon coding agents, whose feedback mechanisms are often ineffective or even detrimental and lack persistent tracking of correction outcomes. To this end, we propose Opera, a framework that introduces a novel "persistent notes" mechanism to transform corrections into traceable, structured records. By integrating a test-time critic architecture, event-driven triggers, typed diagnostic operators, and self-critique strategies, Opera establishes an end-to-end closed-loop monitoring pipeline from feedback generation to outcome verification, effectively distinguishing superficial compliance from substantive resolution. Experimental results demonstrate that Opera improves resolution rates by up to 15 percentage points across multiple benchmarks. Furthermore, after fine-tuning, the model achieves performance comparable to stronger baselines while maintaining robustness.

Abstract

Long-horizon coding agents need timely corrections, yet feedback can be ineffective or even harmful when it misjudges ongoing work or fails to address the underlying problem. Existing critics focus on evaluating trajectories and generating feedback, but rarely track what happens after feedback is delivered. We present Opera, a verbal critic framework that treats each correction as a persistent note, followed until the diagnosed problem is resolved. Opera decides when to review through periodic and event-driven triggers, diagnoses issues with typed operators, audits feedback against visible evidence before delivery, and tracks the agent's subsequent actions to distinguish mere compliance from actual resolution. As a test-time critic, Opera improves the resolve rate of non-critic agents by up to 12.4, 15.0, and 8.9 percentage points on Terminal-Bench 2.1, a SWE-Bench Pro subset, and DeepSWE v1.1, respectively, across four policy models, and achieves the highest mean resolve rate among competitive critic baselines on all three benchmarks, and also improves policy models when the policy critiques itself. Beyond inference, Opera-guided rollouts provide approximately on-policy training data: fine-tuning Qwen3.5-9B on them improves its resolve rate on held-out SWE-Bench Pro repositories by 10.2 percentage points without a critic at inference time, matching fine-tuning on rollouts from a stronger model, while preserving its performance when switching harness, i.e., from Openhands to Terminus-2, which the latter substantially degrades.

Problem

• long-horizon coding agents
• verbal critic
• feedback tracking
• issue resolution

Innovation

• Verbal Critic Framework
• Long-horizon Coding Agents
• Persistent Note Tracking
• Feedback Auditing
• On-policy Fine-tuning