As Artificial Intelligence evolves from single-turn chat interfaces into autonomous, always-on production systems, AI agents are increasingly operating continuously in the background. In this paradigm shift, tokens are no longer just an ephemeral metric for API billing—they have become fundamental production resources that require systematic organization, governance, and dynamic scheduling.
When autonomous agents run non-stop and model capabilities and pricing fluctuate rapidly, how can organizations allocate intelligent compute under strict budget and quality constraints?
The talk will cover the following aspects.
1) Always-On Agents & Token Workloads: Empirical evidence from real-world agent execution (OpenRouter data, Goldman Sachs industry forecasts), uncovering why multi-step task costs diverge drastically and where tokens are actually spent.
2) Frontier Capabilities & Price Deflation: How the purchasing power of intelligence is shifting—measuring capability progress (ECI index), the commoditization of flagship-grade reasoning (e.g., GPT-4o-level performance), and tiered model economics.
3) Jev & The Decision Model Ecosystem: Experiencing the Jevons Paradox in AI—how ultra-low-cost, low-latency decision models (such as Jev) decouple rapid classification, routing, and judgment from heavy generative LLMs.
4) Task-Level Intelligent Scheduling: The frontier where the scheduling layer itself becomes modeled—featuring dynamic routing (RouteFM), feedback payback economics (SaveRouter), collaborative multi-model consensus (Fusion & Fugu), context and KV-cache reuse (C²KV & RaReCache), and trajectory-based token forecasting (TokenCast).
5) Engineering Practice & Token Governance: Actionable blueprints for structuring multi-model pipelines, leveraging low-cost pricing windows, and transforming volatile token consumption into verified, reusable enterprise assets (codebases, test suites, and persistent knowledge bases).

Haofen Wang is a tenured professor and doctoral supervisor at Tongji University. His research focuses on knowledge graphs, knowledge-enhanced large language models, and industrial AI applications. His key contributions include KAG-Thinker, the world’s first knowledge-augmented reasoning model; MemOS, a memory operating system; and AI-Ceping, a large language model evaluation platform, with applications in airport operations and financial risk management. He has authored six books and published over 100 papers, receiving more than 16,000 citations on Google Scholar. He has led more than ten national AI research projects, including key projects funded by the National Natural Science Foundation of China and research tasks under the New Generation Artificial Intelligence Major Project. Previously, he served as CTO of two AI unicorns, whose intelligent customer service systems have collectively handled over 10 billion service interactions. He currently serves as the inaugural rotating chair of OpenKG, the world’s largest open knowledge graph community; deputy chair of the Terminology Committee and secretary-general of the Technical Committee on Natural Language Processing of the China Computer Federation (CCF); a council member of the Chinese Information Processing Society of China (CIPS), an executive committee member of its Technical Committee on Large Language Models, and deputy secretary-general of its Technical Committee on Language and Knowledge Computing; and deputy chair of the Technical Committee on Natural Language Processing of the Shanghai Computer Society.