Help Without Being Asked: A Deployed Proactive Agent System for On-Call Support with Continuous Self-Improvement

Researchers at ByteDance have introduced 'Vigil,' a novel AI agent system designed to transform technical support by operating proactively alongside human analysts. Deployed on ByteDance's Volcano Engine cloud platform for over ten months, Vigil addresses a critical gap left by first-line reactive chatbots. When these initial bots fail to resolve an issue and a human analyst is escalated, traditional agents disengage. Vigil, however, stays in the conversation, monitoring the dialogue between the customer and the human to offer relevant suggestions, track progress, and provide assistance without needing to be explicitly invoked.

The core innovation is Vigil's dual capability: proactive intervention and autonomous learning. The system uses large language models (LLMs) to understand the context of live support tickets and interject with helpful information or potential solutions. More importantly, it features a continuous self-improvement loop. After a case is closed by a human, Vigil analyzes the resolution to extract new knowledge, which it uses to autonomously update its own internal reasoning and response capabilities. This creates a system that gets smarter with every ticket handled, moving beyond static scripted responses.

The paper presents a comprehensive evaluation based on the real-world deployment, showing Vigil's effectiveness in a high-volume environment where thousands of tickets are generated daily. By assisting during the most labor-intensive phase of support—the human-in-the-loop escalation—Vigil directly reduces the cognitive load and time pressure on support staff. The researchers have also made an open-source version of the work publicly available, encouraging further development in the field of proactive, collaborative AI agents for enterprise applications.