In the AI-Native Era, Let the World Adapt to Agents, Not Teach AI to Be Human | HKU Huang Chao @ AIGC2026

In the rapidly evolving landscape of AI agents, a fundamental question arises: should we continue teaching AI to use human tools, or should we redesign the digital world to speak the language of machines? At the 2026 China AIGC Industry Summit, Professor Huang Chao from the University of Hong Kong made a compelling case for the latter. His keynote, "Make the World Adapt to Agents, Not Teach AI to Be Human," challenged the prevailing approach of making AI mimic human interfaces and proposed a radical shift toward agent-native infrastructure.

Huang's team began by questioning the complexity of existing agent frameworks. When they observed that OpenClaw had 430,000 lines of code (and has since exceeded one million), they chose the opposite path: building an ultra-lightweight, open-source general-purpose agent called nanobot. The response exceeded expectations. Nanobot was updated daily for 100 consecutive days, accumulated over 200,000 downloads, was selected by DeepSeek as one of 15 globally recommended agents, and briefly ranked fourth on the OpenRouter general agent leaderboard. This success underlines a philosophy of elegance: the core of agent technology does not require overwhelming complexity.

But lightness is just the beginning. Huang believes the next challenge is upgrading agents from simple AI assistants to true productivity tools. To achieve this, they propose CLI-Anything — a system that repackages professional software into command-line interfaces. Instead of forcing agents to navigate graphical user interfaces (GUIs) with costly and error-prone visual recognition, CLI-Anything lets agents directly invoke software functions through structured commands. Huang argues that CLI is the true AI-native way of computer use. "Why make AI click like a human when it can communicate more precisely through a command line?" The CLI Hub already hosts 80 software tools across 31 categories. The vision is a hybrid future: CLI for agents, GUI for humans — each using its most natural interaction mode.

Beyond interfaces, Huang addressed agent self-evolution. His research identifies three paths: internal optimization (tuning the agent's core), parameter updating (learning through reinforcement), and skill accumulation (collecting reusable skills). The internal and parameter approaches suffer from poor generalization — an agent optimized for one task struggles to transfer. The team therefore focuses on skill accumulation via a system called Open Space. However, challenges remain: high-quality skills are scarce, matching skills to tasks is non-trivial, and managing skills at different granularities is complex. Their experiments across 220 tasks from 44 industries showed significant reductions in token consumption and improved task completion after agents learned from real tasks.

Huang also explored agent swarm collaboration. In a striking experiment, eight agents coordinated eight H100 GPUs to train a large language model. Within 23 hours, model performance improved by 6%, equivalent to three weeks of work by a PhD student. This confirms the potential of multi-agent systems, but also reveals a critical insight: the scaling law for agent swarms is not straightforward. Beyond 3–5 agents, coordination overhead can exceed performance gains. Zero communication delay does not automatically unlock larger teams; the bottlenecks shift to task decomposition, conflict resolution, and quality validation. Finding the optimal swarm size remains an open challenge.

Huang's overarching message is clear: the future of AI agents is not about making them more human, but about building an ecosystem where agents can operate natively. From the lightweight nanobot to CLI-Anything to skill-based evolution and swarm experiments, each piece points toward a coordinated infrastructure designed for agents. As he concluded, "The value of future agents lies not in individual breakthroughs, but in the synergy of the entire ecosystem."