跪拜 Guibai
← All articles
Algorithms · LLM · Agent

From Rigid Scripts to Service Partners: Deploying a Controllable Multi-Agent Customer Service System

By 得物技术 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most agent demos collapse under real customer-service loads: long contexts, multi-turn negotiations, and the need to sound human while staying on-policy. This stack shows a repeatable path from a naive LLM call to a trained, evaluated, and controllable multi-agent system, with concrete numbers on resolution rates and alignment with top human performers.

Summary

Traditional customer-service bots break on multi-step tasks and sound robotic. Dewu's team replaced a single-prompt Qwen agent with a multi-agent architecture built on AutoGen, using a master controller, response, evaluator, and polishing agent to handle complex negotiations. The system now runs under a cross-scenario Harness that adds cross-session memory, context compression, and a data flywheel for continuous improvement.

A prompt-engineering automation pipeline compares bad agent replies against good human replies, proposes prompt changes, and validates them offline before online experiments. On the training side, DPO and GRPO-based RL teach the model correct decisions, with specific countermeasures against reward hacking and catastrophic forgetting. A separate distillation pipeline aligns the agent's tone, emoji use, and decision logic with the top 5% of human agents.

Half-duplex message control handles the reality that users can fire off multiple messages while the agent is still generating a reply. A dedicated Qwen3-4B classifier decides when to merge or split user messages, keeping the conversation coherent without dropping context.

Takeaways
A single-prompt agent fails on multi-step tasks; a multi-agent setup with a master controller, response agent, evaluator, and polisher raised resolution rates significantly.
Prompt engineering is automated by diffing bad agent replies against good human replies, having an LLM propose prompt edits, and validating them on offline data before pushing to production.
DPO training uses LLM-as-a-Judge to build discriminators for reply-ability, correctness, and human-likeness, then constructs chosen/rejected pairs for the polishing model.
GRPO-based RL trains decision-making, with rule-based and model-based rewards; catastrophic forgetting is held off by low learning rates, early stopping at 2–4 epochs, and a 1:1 ratio of self-reasoning to tool-call data.
Human tone and tactics are distilled via CoT-RL for decision logic and DPO for dialogue style, including a 12-class emoji classifier that controls when and which emojis appear.
A half-duplex message protocol lets users interrupt freely but prevents the agent from being cut off mid-reply; a Qwen3-4B classifier decides whether consecutive user messages should be merged or split.
Conclusions

Prompt engineering, often treated as a craft, is reframed here as a data-flywheel problem: the system proposes prompt edits from outcome diffs, and humans only review, which turns prompt tuning from a bottleneck into a continuous loop.

Training separate discriminators for reply-ability, correctness, and human-likeness, then using them to construct DPO pairs, is a practical decomposition that avoids trying to capture all quality dimensions in a single reward signal.

The half-duplex design acknowledges a real UX constraint most chatbot frameworks ignore: the agent's inference latency means user messages pile up, and a dedicated classifier for merging or splitting them is cheaper and more reliable than trying to handle it inside the main agent prompt.

Concepts & terms
Multi-Agent Architecture (AutoGen SelectorGroupChat)
A setup where multiple specialized agents (master controller, response, evaluator, polisher) collaborate through dynamic message routing, rather than a single monolithic prompt handling the entire conversation.
PE Automation Pipeline
A loop that extracts bad agent replies and good human replies from escalated cases, uses an LLM to suggest prompt changes, has humans review them, and then validates the changes on offline data before online deployment.
GRPO (Group Relative Policy Optimization)
A reinforcement-learning method used here to train the agent's decision-making, combining rule-based verifiable rewards with model-based generative rewards, while adding chain-of-thought reasoning to counter reward hacking.
Half-Duplex Message Control
A communication mode where users can send messages at any time (full-duplex), but the agent's outgoing replies are protected from interruption; a classifier model decides whether consecutive user messages should be merged or treated as separate turns.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗