Interview: Xia Lixue on Building Agentic Infrastructure for Large-Scale AI Deployment

Edited by: Luo Yanshan, Tina

Curated by: Yuqi

---

Introduction: From AI Infra to Agentic Infra

A new era—driven by intelligent agents as the fundamental operational unit—is accelerating. Infrastructure is evolving from AI Infra → Agent Infra → Agentic Infra, becoming the critical force enabling large-scale agent deployment.

As one of China’s leading artificial intelligence infrastructure service providers, Wuwen Xinqiong is tackling the challenges of agent collaboration, security, and continuous learning—transforming them into real productive assets.

From October 23–25 in Shanghai, during the QCon Global Software Development Conference, InfoQ spoke onsite with Xia Lixue, co-founder and CEO of Wuwen Xinqiong, about how Agentic Infra can support agent scaling and industrial deployment.

---

Key Insights from Xia Lixue

• Diverse, distributed computing resources in China demand infrastructure that lets agents access compute as easily as water or electricity.
• Agent Infra must evolve from a “production line factory” to a “solution company,” ensuring task execution quality in environment, tools, context, and security.
• Advancing to Agentic Infra requires greater autonomy—enabling more efficient resource integration and functional innovation at higher value.
• Smarter infrastructure creates new demands → demands drive R&D upgrades → a dual flywheel between technology and application.

---

Paradigm Shift: From “Processing” to “Thinking”

The Agentic Loop

Traditional cloud computing runs on deterministic request–response models. In contrast, Agentic AI operates in a perception → reasoning → action → memory loop—a non-linear, stateful cognitive process.

> Xia Lixue:

> Agents' tasks are correlated, continuous systems—not discrete jobs. Infrastructure must acquire intelligence to guarantee the overall output quality of agents.

Necessary upgrades include:

• Adaptive runtime environments to match agent execution modes.
• Robust tools for agents to utilize resources effectively.
• Complete context information for task understanding and execution consistency.
• Safety and monitoring mechanisms for controllability and observability.

Key takeaway: Agents need longer “thinking” time and new kinds of resources → requiring flexible allocation strategies, real-time scheduling, and design logic rethink.

---

Scaling Challenges & Market Signals

• Example: CrustData reports Lovable platform users fell 40% from June to September.
• Cause: High public expectations for “no-code programming” vs. reality—Vibe Coding still requires iteration and domain knowledge.
• Core bottleneck: Immature infrastructure services/tools, not weak models.
• Solution:
• Process control and observability—clear changes, decision paths.
• Tool accessibility—prevent “guessing” due to missing permissions or APIs.

---

From Agent Infra → Agentic Infra

Stage 1: Agent Infra

• Push agents out of labs into productivity scenarios.
• Handle multi-agent collaboration, resource competition, task conflicts, cost control.

Stage 2: Agentic Infra

• Agents deeply participate in infrastructure workflows.
• Systems autonomously detect anomalies, locate bottlenecks, optimize resources, and scale elastically.
• Shift mindset from agents as tools → agents as collaborators.
• Build frameworks for efficient, low-cost agent cooperation and evolution.

---

Why Now Is the Right Moment

Even without a clear generational leap in base models in the last 6 months, real-world use cases keep expanding. Models are “smart enough”—but lack the right environment.

---

The Path Toward Autonomous Infrastructure

Analogy: A genius programmer shouldn’t start as CTO—gradual skill-building applies to agents. Today, humans predefine workflows; Agentic Infra seeks agents designing their own.

A2A (Agent-to-Agent) Collaboration:

• Agents autonomously learn, collaborate, and divide tasks.
• Signs include backend operations agents + frontend customer service agents outperforming human relay methods.
• Build Infra to support swarms of intelligent agents without bottlenecks.

---

Autonomous Behavior in Infrastructure

> Xia Lixue: Once Infra integrates Agent capabilities, autonomy emerges—similar to cross-department collaboration in organizations.

> Breaking “role boundaries” boosts value via efficient resource integration and functional innovation.

---

Future Vision: Agent-Native Architecture

3–5 Years Ahead:

• Agents form organizations to achieve complex tasks.
• Shared/isolated KV Cache & context per task needs.
• Outcome: Free human mental + productive capacity for high-value creative work.

---

Computing Power Allocation: Breaking Coarse-Grained Limits

Problems with Traditional AI Infra:

• Fixed-size virtual machines/containers ill-suited for small, parallel agent tasks.
• Cold starts and reserved lock scheduling waste time and resources.

Solution:

• Micro-virtualized sandboxes + high-concurrency management.
• Dynamic capacity assembly, millisecond-level switching, ~100% utilization.

---

Unified Scheduling of Heterogeneous Compute Power

Innovation:

• Resource standardization across architectures → efficient task allocation.
• In China’s fragmented compute ecosystem, agents must tap resources as easily as utilities.
• Sandbox system supports heterogeneous compute, leveraging accumulated technology.

---

From Technical Breakthrough to Industrial Integration

• Early stage: Technical adaptation—cross-chip communication, operator unification.
• Post-breakthrough: Focus on usability—hide low-level differences from users.
• Agent era: User cares about task completion quality, not chip brand.

---

Balancing Tech Advancement & Customer Service

> Customer pain point: Static sandboxes without elastic scaling → resource waste or shortage.

> Response: Elastic scaling + dynamic scheduling sandbox system → automatic CPU/GPU/chip mounting, millisecond-level start/stop, workload-based allocation.

Result:

• Addresses user needs and advances Agentic Infra tech.
• Technology ↔ application form a positive dual flywheel.

---

Personal Mission & Vision

Xia Lixue:

• Witness AGI era arrival; accelerate progress with system-level innovation.
• Vision: “Unleash limitless computing power and make AGI within reach.”

---

Ecosystem Parallel: AiToEarn

Platforms like AiToEarn官网 mirror Agentic Infra principles:

• Open-source AI content monetization.
• Cross-platform publishing to Douyin, Kwai, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X.
• Integrated content creation, analytics, model ranking.

---

Closing Thought

Bridging robust infrastructure with application ecosystems—whether for AI agents or human creators—is vital. Unified scheduling maximizes compute efficiency; unified publishing maximizes creative monetization. Both are stepping stones toward the Agent-native future.

---

Would you like me to also prepare a visual summary diagram that maps "AI Infra → Agent Infra → Agentic Infra" stages with key capabilities? That could make this piece even more reader-friendly.