The AI Support Agent That Reads the Room — Without Being Asked

ByteDance deployed an LLM agent that silently monitors live customer support chats and jumps in before the human analyst gets to it — and it keeps getting smarter from every ticket it watches.

~5 min read · ByteDance / Volcano Engine · 2026-04-14 · Agentic

TL;DR Most LLM-based support agents bail out the moment a human takes over a ticket. Vigil does the opposite: it activates when the human joins, monitors the live conversation, proactively answers sub-questions the analyst hasn't gotten to yet, and then mines every resolved case to improve itself — all without anyone pressing a button. Deployed on ByteDance's cloud platform, it has handled over 131,000 on-calls across nearly 2.3 million messages.

There's a maddening gap in how AI agents are typically deployed for customer support. An LLM bot handles the easy stuff, fails on the hard stuff, and then — the moment a human expert joins — politely exits the room. The human now has to read the full conversation history, understand what was already tried, and respond in real time to a customer who's already frustrated. The AI that could help is gone.

Vigil is built on the observation that this handoff moment is precisely when AI assistance is most valuable, not least.

💡 The Core Idea Instead of replacing the first line of support, Vigil acts as a silent co-pilot during the human-involved phase of a support session. It watches the conversation, identifies customer sub-questions the analyst hasn't answered yet, and proactively inserts answers — no one has to summon it.

Think of it like a junior analyst sitting next to a senior one, whispering "the customer also asked about X three messages ago — here's what I found" before the senior has a chance to scroll back up.

How Vigil Works

Vigil's action cycle starts the moment a human analyst enters an on-call session and ends when the session closes. During that window, it runs two parallel processes.

flowchart LR
    A[Customer Message] --> B[Question Identification]
    B -->|Valid unanswered sub-question| C[Answer Generation]
    B -->|Already addressed or out of scope| D[Stay Silent]
    C --> E[Duplicate Check]
    E -->|Novel answer| F[Proactive Response]
    E -->|Duplicate| D
    F --> G[Self-Improvement Engine]
    G -->|Extracts knowledge| H[Knowledge Base]
    H --> B

Online Proactive Response involves three steps. First, a context-aware question identification module decides when to intervene — specifically targeting moments when a customer raises a valid sub-question that the analyst hasn't yet addressed. This is the "when before what" problem: jumping in at the wrong moment creates noise, not help. Second, Vigil generates an answer. Third, a duplicate-prevention layer checks whether Vigil has already offered a similar answer earlier in the same session, avoiding the AI equivalent of a broken record.

Continuous Self-Improvement is the other half of the system. Rather than sitting static after deployment, Vigil reviews resolved cases and extracts new knowledge entries autonomously. Every ticket a human successfully closes becomes training signal — Vigil mines it and updates its own knowledge base without human curation.

What a Deployment at Scale Looks Like

Vigil went live on Volcano Engine — ByteDance's public cloud platform — on March 13, 2025, starting with a limited set of on-calls before expanding to full coverage by June 1, 2025.

By January 16, 2026, the numbers tell the story of real production load:

That's a system that has been quietly reading millions of support messages and choosing, on its own judgment, when to speak up — over ten months of continuous operation.

Why the Reactive Agent Paradigm Falls Short

The standard approach to LLM-based on-call support treats the AI as a first-line filter: it tries to resolve tickets, and anything it can't handle gets escalated to humans. At that point, the AI is done. This design has two compounding problems.

First, customers rarely arrive with a single, clean, fully-formed question. A cloud service issue typically unfolds: the customer describes a symptom, the analyst investigates, the customer asks a follow-up, something unexpected appears in the logs. A reactive agent that already bowed out can't track any of this evolution.

Second, the hard cases — the ones humans have to resolve — are the most knowledge-rich cases for improving the system. Reactive agents miss all of that signal entirely, because they're not in the room when the resolution happens.

Vigil addresses both gaps simultaneously. By staying engaged through the human-involved phase, it can respond to follow-up questions in context. And by watching how human analysts actually resolve difficult tickets, it can extract the reasoning and solutions that it couldn't generate on its own.

⚠️ Watch out for

• The paper does not detail how Vigil handles incorrect proactive responses — in a live customer dialogue, a wrong confident answer from the AI could undermine analyst credibility
• Multi-modal inputs (screenshots of error logs, monitoring dashboards) are noted as part of real on-call inquiries, but the full extent of Vigil's visual understanding isn't described in the excerpted sections
• Self-improvement via automated knowledge extraction could propagate errors if a human resolution itself was suboptimal

Why It Matters

The cloud support context is a useful stress test for proactive agents because the stakes are high (customers are down, SLAs are ticking) and the volume is enormous. If a system can work here — silently, correctly, without being invoked — it can work in a lot of other human-in-the-loop workflows where a knowledgeable AI co-pilot would save time without taking over.

The self-improvement loop is arguably the more consequential architectural choice. It transforms every escalation from a failure event into a learning event. A system that gets smarter from the cases it couldn't handle is a fundamentally different trajectory than one that requires periodic manual retraining.

The code is publicly available, which means the architecture can be adapted beyond cloud support to any domain where human experts handle high-volume, complex, sequential dialogues.

Source: Help Without Being Asked: A Deployed Proactive Agent System for On-Call Support with Continuous Self-Improvement Authors: Fengrui Liu, Xiao He, Tieying Zhang Published: 2026-04-14 PDF: https://arxiv.org/pdf/2604.09579