How Agents Are Becoming the New Employees in Enterprises – Insights from AIEC 2026

In just a few months the large‑model landscape has pivoted from the traditional "bigger, stronger, more general" mantra to a focus on powerful coding capabilities, and now to the concept of "Harness" – a complete engineering system that straps a strong AI model with a "reins" so it can operate reliably and efficiently within real‑world enterprises.

The AI+Eco Conference (AIEC 2026) in Beijing, organized by Tsinghua University's Global Industry Research Institute and supported by the Zhongguancun Science City Management Committee, gathered experts from national information centers, IDC, Alibaba Cloud, Tencent, and many leading AI labs to discuss how AI moves from strategic deployment to practical industry practice.

Speakers highlighted that programming scenarios already consume over 90% of model token usage, yet there are only 30 million programmers worldwide compared with more than a billion knowledge workers, indicating a massive upcoming expansion of AI assistance beyond code.

Peng Zhen, chairman of Inspur, argued that AI changes not just tools but the labor force itself, reshaping production relations and requiring organizations to manage "Human+Agent" (Humagent) rather than just humans.

Model companies are now building their own coding tools and agent frameworks because a pure model no longer delivers enterprise value; the real commercial impact lies in agents that can directly complete tasks, manage context, invoke tools, recover from errors, and collaborate with other agents.

Wang Shengjie, senior product‑technology expert at Tencent, emphasized that scaling enterprise agents requires a unified Agent OS and intelligent scheduling to turn agents into a "super team" capable of management, execution, verification, and memory.

Han Hongna, solution director of Alibaba Cloud Qoder, described the evolution of AI coding tools in three stages: code completion, autonomous development workbenches, and finally expert‑agent teams that handle end‑to‑end tasks.

Chen Yutao introduced the MemTensor solution, breaking agent memory into five capabilities—extraction, organization, retrieval, update, and sharing—and proposing "Memory Skill" to capture expert experience as reusable skills for other agents or human training.

Gong Guan identified three practical bottlenecks for agents in production: latency (affecting user retention), cost (high token consumption for long‑running inference), and reliability (risk of task drift or interruption). The industry’s next frontier is not extreme intelligence but scalable, high‑efficiency intelligence.

The competition timeline outlined by speakers shows that from 2023‑2025 the battle was on the model layer (parameter count, benchmark scores), in the first half of 2026 it shifted to the tool layer (agents and coding frameworks), and from the second half of 2026 it will move to the organizational layer, where firms that help enterprises restructure around AI agents will gain the strategic advantage.

Finally, Professor Peng Kaiping warned that while AI boosts efficiency, human competitive advantage lies in psychological abilities—empathy, moral judgment, creative imagination, and intuitive decision‑making—calling for a "mind sovereignty" that preserves human values, privacy, and agency in the AI era.