CTO Argues for Building 'Systems,' Not 'Features,' with AI
Adam Bilišič, CEO of Nodeon Labs, argued that engineering teams using AI-assisted coding should shift from building features to creating reusable "systems" or "recipes." He suggests that competitive advantage now lies in creating robust, modular sub-agents and libraries of reusable prompts that encode best practices and decision logic. Bilišič stated, "AI coding tools are not meant to replace developers, but to augment their capabilities."
- The shift from features to systems often involves adopting multi-agent orchestration frameworks like LangGraph or Microsoft's Agent Framework, which use graph-based structures to manage complex workflows between specialized agents. Google's Agent Development Kit (ADK) offers eight specific design patterns for this, including sequential pipelines for predictable tasks and parallel fan-out/gather patterns for simultaneous operations, helping to improve modularity and reliability. - A primary challenge in multi-agent systems is managing "handoffs" between agents, as failures often occur due to inconsistent state synchronization or accumulated latency from coordination overhead. Production systems show that a workflow with 10 agent handoffs can add 1-5 seconds of pure coordination overhead before any processing occurs, making explicit, compressed handoff payloads critical for scalability. - As AI engineering teams scale, communication overhead and unclear ownership become critical failure points. Frameworks like Team Topologies are being adapted for AI, organizing teams around specific product areas ("stream-aligned") with a "You Build It, You Run It" mindset to maintain clarity and control as system complexity grows. - In Beijing, the competitive landscape for AI agents is accelerating, with Alibaba's DingTalk launching a marketplace of over 200 AI agents and the government actively supporting local startups like Manus. Additionally, Beijing's Chaoyang district recently launched an AI agent innovation accelerator to support over 30 startups, signaling a strong local ecosystem focus. - For consumer-facing AI agents, user experience is shifting toward iterative refinement and control. Key AI interaction patterns include providing suggested prompts to reduce "blank page" paralysis, offering inline actions to edit or regenerate output directly, and implementing user feedback loops that allow the system to learn and personalize its responses over time. - China's regulatory environment for AI is rapidly maturing. An amended Cybersecurity Law took effect on January 1, 2026, which now includes AI governance and raises the maximum fine for non-compliance to CNY 50 million or 5% of the previous year's turnover. Furthermore, draft measures issued in late 2025 specifically target anthropomorphic AI, proposing strict rules on content, data protection, and preventing user addiction. - Research in multi-agent systems is increasingly focused on emergent coordination and dynamic role allocation, where agents can autonomously assume roles based on their capabilities and system needs. Studies show that multi-agent reasoning frameworks can improve performance on complex logical tasks by over 21% compared to single models. - Architecturally, a common approach for enterprise-grade multi-agent systems is the "supervisor pattern," where a primary agent orchestrates and dispatches tasks to a collection of specialized agents. This modular design, seen in platforms like Databricks' Mosaic AI, allows different business divisions to develop and manage their own agents independently while ensuring central coordination.
