Help Without Being Asked: A Deployed Proactive Agent System for On-Call Support with Continuous Self-Improvement
Vigil is a rare example of a proactive agent deployed at production scale (ByteDance Volcano Engine, 10+ months). It covers the full on-call lifecycle — including the post-escalation phase that reactive agents abandon — and continuously learns from human-resolved cases, making it a strong reference for real-world proactive system design.
Vigil integrates into human–customer dialogues and proactively offers assistance without explicit invocation. It monitors ongoing cases, identifies resolution gaps, and surfaces relevant knowledge before the analyst asks.
Evaluated on real deployment data from Volcano Engine. Metrics cover assistance quality, analyst workload reduction, and self-improvement efficacy over the 10-month deployment window.
Evaluation is on a single internal cloud platform; generalization to other domains is not measured. The continuous self-improvement pipeline depends on the quality of human-resolved case logs.
Use for real-world proactive agent deployment, on-call/IT-support automation, continuous self-improvement mechanisms, and deployment-scale evaluation studies.