From Answering Questions to Completing Tasks: What Do Consumer Agents Still Lack? Fliggy's New AI Product V10 Attempt

Fliggy's new AI product V10 is a attempt to solve the challenge of consumer agents moving from answering questions to completing tasks. The CTO Chen Ye noted that value in AI is still concentrated in base models, GPUs, and infrastructure, but should extend to application and orchestration layers. V10 introduces an AI Panel integrated into pages, with capabilities '帮我想' (idea), '帮我订' (booking), and '帮我办' (service). The algorithm architecture evolved from Multi-Agent with workflow to ReAct-style AgentLoop, but with 'constrained autonomy': using short aliases for products, whitelist mechanisms, rules, and light verification to ensure real supply. The team uses a three-stage approach: Prompt (declarative), mechanisms, and training, with a reward model covering truthfulness, user needs, spatiotemporal validity, and completeness. They balance LUI and GUI, using LUI for complex long-tail needs and GUI for efficient browsing and trading. Evaluation focuses on usability rate, with multi-model voting and human review. The article concludes that vertical applications like travel may provide new training environments for AI, but migration from coding to consumer agents will take time, maybe two years or more, as it requires integrating real-world data, evaluation, and supply chain.