How Dual Forums at Agentic AICon Shanghai Redefined AI‑Agent Integration for Real‑World Impact

Conference Core Theme

The event featured two parallel flagship forums with clear focus: AI for SE (AI empowering software engineering) and SE for AI (software engineering empowering AI agents), forming a complete content matrix.

Industry Context and Challenges

The AI industry has moved beyond single‑model pilots; competition now centers on deployment capability, engineering systems, and product commercialization. Rapid iteration of large models and AI agents creates industry trends of mutual empowerment between AI and software engineering, yet faces difficulties such as complex system adaptation, hindered human‑machine collaboration, lack of standardized design methods, insufficient architectural support, and distorted evaluation metrics.

Keynote 1 – "AI Agents Empowering Software Development in the Real World" (AI for SE)

Speaker Zhang Gang (Senior Technical Expert, Shanghai Yingmu Technology) presented the evolution‑oriented design for AI agents, arguing that static software design patterns cannot accommodate the dynamic nature of agents. He detailed four core directions: agent‑centric architecture paradigm shifts, boundary analysis of complex systems, mass‑production‑grade AI‑enabled development pipelines, and engineering‑grade evaluation standards. The talk addressed agent iteration in complex scenarios, human‑machine division of labor, organization‑level productivity gains, and quantitative assessment of real‑world engineering capabilities, constructing a complete, deployable AI‑for‑SE engineering loop.

Keynote 2 – "The Beauty of Agent Design Patterns: Building an Open, Self‑Evolving Agent Engineering Paradigm" (SE for AI)

Speaker Peng Xin (Associate Dean, Fudan University; CodeWisdom Team Lead) examined software‑centric perspectives on AI agents, using metaphors like “a horse crossing a river” to illustrate common misconceptions about project complexity and development models. He distinguished programming from software development, applied Brooks' essential and accidental complexity theory, and analyzed the operational characteristics of complex software systems. By referencing OpenAI coding agents, he defined the boundary of AI capabilities, promoted the principle "AI is a chief of staff, not a commander," and shared team explorations in code knowledge extraction, feature‑driven iteration, and constraint‑aware AI development.

Additional Keynotes

Chen Xin (Alibaba Qoder IDE Lead) traced the evolution from chatbots to desktop IDEs, dissected three generations of prompting, context, and harness engineering, compared major multi‑agent architectures, and revealed Alibaba's self‑developed equal‑collaboration agent framework, layered knowledge engine, companion memory agents, and three‑layer security sandbox.

Ru Bingcheng (Tencent Research Institute) critiqued current AI coding evaluation benchmarks, identified eight systemic flaws (simplified scenarios, idealized environments, biased metrics, data contamination, missing token cost, etc.), and proposed a three‑layer ideal evaluation architecture: programming cognition → engineering practice → human‑machine collaboration.

Wang Haofeng (Tongji University) presented memory system construction for agents, comparing internal‑vs‑external memory routes, detailing lifecycle stages (identification, construction, retrieval, update), and showcasing open‑source MOS project and its iterations, while discussing privacy, hallucination, and long‑video processing challenges.

Huang Jia (A*STAR AI Researcher) argued that traditional OOP patterns are unsuitable for AI agents, introduced a dual‑axis framework (seven execution topologies, seven cognitive functions), and announced the launch of the ADPS (Agent Design Patterns Society) open‑source project to unify design standards.

Deng Yafeng (EverMind CEO) outlined next‑generation Agent OS concepts, compared external storage, parameter storage, and hidden‑state long‑term memory solutions, and detailed Evermind's IOS framework and LME personal memory center.

Yan Junchen highlighted the transition from single agents to a full Harness stack, covering organizational, product, and commercial dimensions, and offered practical steps for scaling agent stability and iteration.

Round‑Table Discussion

Experts from academia and industry debated the value and pitfalls of AI coding tools, focusing on agent capability boundaries, complex‑scenario deployment, human‑machine division of labor, and risk mitigation, providing actionable insights for practitioners.

Conference Significance

Addressed practical deployment challenges by covering AI‑enabled software engineering and software‑enabled AI agents.

Bridged academia and industry through collaborations with top universities, research institutes, and leading tech companies.

Launched the ADPS open‑source design‑pattern project to standardize agent development.

Showcased transformative case studies that illustrate how AI reshapes traditional software development.

Delivered high‑quality knowledge resources, including multiple new books on agent design patterns.

Overall Outcome

The conference delivered extensive technical solutions, real‑world case studies, and evaluation methods, while officially initiating the ADPS open‑source initiative to drive standardization, engineering, and self‑evolution of AI agents in China. It established a robust platform for industry‑wide collaboration and set a clear direction for the future of AI‑agent‑driven software engineering.