Is 2025 the Dawn of the AI Agent Era? Expert Insights from the Inclusion Conference
At the 2025 Inclusion·外滩大会 forum, leading academics and industry pioneers discussed rapid advances in AI agents, highlighting breakthroughs in multi‑agent systems, reinforcement learning, open‑source frameworks, and the practical challenges of cost, performance, and usability that still separate "usable" from truly "useful" technology.
2025 has been dubbed the "Year of the Agent" as large‑model technology accelerates and agent applications emerge, prompting a debate on whether an AI‑driven transformation is truly underway.
On September 11, the Inclusion·外滩大会 "Intelligent Agent Evolution Theory" forum gathered scholars and industry leaders to discuss breakthroughs, multi‑agent collaboration, open‑source ecosystems, and real‑world deployment challenges.
Significant technical breakthroughs: multi‑agent systems and reinforcement learning drive the momentum
Prof. Liu Zhiyuan (Tsinghua University) highlighted the transition from large models to agents, emphasizing core challenges of generalization, autonomy, and long‑term reasoning, and described multi‑agent systems as the infrastructure for next‑generation collective intelligence.
Zhu Zhiqing (Founder, Pokee AI) explained how reinforcement learning combined with large language models—through RLHF and RLVR—dramatically improves performance in mathematics and code generation, citing DeepMind's Gemini "Deep Think" achieving gold‑level scores on IMO problems.
Wu Chenglin (CEO, DeepWisdom) introduced MetaGPT, an open‑source multi‑agent collaboration framework with over 150 k GitHub stars, and demonstrated rapid prototyping of OpenManus using the framework.
Fan Wendong (Core contributor, CAMEL‑AI) showcased the project's wide applications in task automation and social simulation.
Gu Jinjie (Project Lead, Ant Group InclusionAI AWorld) described AWorld as a next‑generation framework designed for large‑scale self‑improving agents, enabling autonomous evolution through knowledge integration.
Implementation challenges remain: the gap between "usable" and "useful"
Sheng Sixiong (CEO, Ponder) illustrated the fragmented output problem of current agents using AI‑generated PPTs, arguing for human‑centered, process‑oriented interaction design.
Cost and performance also pose bottlenecks; Zhu Zhiqing noted high inference costs and limited scalability hinder large‑scale commercial deployment, especially under high concurrency.
A roundtable titled "Post‑2000 Agent Startup Industry Observation" featured young entrepreneurs sharing perspectives on the agent ecosystem.
Overall, while agents are poised to become a dominant technology direction in 2025, experts agree they currently serve best as assistants rather than replacements, and bridging the gap from "usable" to truly "good" will determine their transformative impact.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.