On April 18, DAS-Security hosted a Tech Talk at Anheng Building in Hangzhou, featuring Hao Wu from Goofish, who shared practical insights on AI-driven engineering for team collaboration and productivity.

The key topics covered are:

• Integrated Development
• Knowledge-Driven Intelligence
• Multi-Agent AI Collaboration

Rather than revisiting the familiar debate — “Can AI write good code?” — the session focused on a more critical question:

What does it take for AI to truly participate in real-world delivery?

The answer, as highlighted throughout the talk, lies not in the model capability alone. AI has already demonstrated strong capabilities in generating code. However, production environments demand far more than generation — they require reliability, validation, and repeatability.

He highlighted a dilemma nowadays, as addressed by Conway's Law, that any organization that designs a system will produce a design whose structure is a copy of the organization's communication structure. However, this design is likely to be an inefficient one. Essentially, in the fast-paced world of professional environments, project execution knowledge sharing is a cornerstone of success. To move from experimentation to impact, AI must be embedded into a complete delivery system.

Four Key Takeaways

1. The Definition of Delivery Is Changing

The benchmark for AI is shifting. It is no longer about whether AI can generate code, but whether it can consistently deliver verifiable results within a structured workflow.

• Model capability is only the starting point
• The end-to-end delivery loop determines real efficiency gains
• Stability and validation matter more than raw generation ability

In other words, “working code” is not enough — reliable outcomes are the real metric.

2. Real-World Use Defines the Boundaries

Practical implementations show that AI Agents can already participate across the full lifecycle:

• Requirements understanding
• Development
• QA and testing
• Release and deployment

However, several constraints still limit scalability:

• Isolated, “single-agent” workflows
• Low validation and acceptance standards
• Slow feedback loops

These bottlenecks highlight a key reality that AI can enter the workflow, but it cannot yet operate effectively without system-level support.

3. Infrastructure Determines the System Ceiling

One of the most important insights from the session was that infrastructure is not optional, yet foundational.

To elevate AI from a tool to an engineering system, organizations must invest in:

• Execution frameworks: runtime environments for agents
• Protocol layers: standardized interaction mechanisms
• Knowledge modeling: structured, reusable context
• Skill orchestration: task distribution and coordination
• Telemetry and feedback loops: continuous learning and optimization

These elements form the backbone of a scalable AI system. Without them, AI remains a point solution rather than a systemic capability.

4. The Goal Is System-Wide Efficiency

Improving individual developer productivity does not automatically translate into overall efficiency gains.

In complex environments, the goal shifts toward system-wide optimization.

This requires a transition from:

• One-man-in-the-loop → where individuals oversee AI outputs
  to
• Experts-in-the-loop → where domain experts guide high-value decisions

AI handles execution at scale, while humans focus on judgment, strategy, and critical validation.

Conclusion

As AI continues to reshape engineering and security operations, the challenge — and opportunity — lies in building systems where humans and AI work together seamlessly.

We thank Hao Wu and all participants for an engaging and thought-provoking session. More discussions and insights to come.