From Code Generation to Self‑Healing Ops: How AI Agents Drive Enterprise Efficiency

The June 11, 2026 livestream titled “AI Agent落地实战：从代码生成到运维自愈，企业级效能革命” gathered three senior practitioners—Yán Qiáozhì (DeepSecurity), Liú Xíngxíng (Ping An Life) and Xióng Wěi (China Mobile Jiutian)—to discuss why large‑model AI agents still struggle in real‑world enterprises.

All three agreed that the primary obstacle is not the raw ability of the model but the surrounding engineering, organizational and risk‑management framework. Liú highlighted three practical pain points when reverse‑engineering legacy systems: massive codebases (millions of lines), missing documentation, and the need to automatically produce PRDs, flowcharts and architecture diagrams. He warned that feeding the entire codebase to a model hits token limits and hallucination problems.

Liú’s proposed solution is to avoid “one‑shot” ingestion and instead perform link analysis, build a code‑graph, and then process the system in “layer‑, domain‑, slice‑” chunks so that each model call handles a focused, controllable context. This turns AI from a direct replacement into a tool that is first “tamed” by engineering methods.

Xióng emphasized that in operations the critical issue is not whether an agent can detect problems but whether its recommendations are trustworthy, auditable and rollback‑able. He argued for a strict separation: hard rules (safety lines, change‑approval, audit trails) must govern high‑risk actions, while the model handles probabilistic tasks such as fault attribution, experience recall and solution suggestion.

Yán focused on the end‑to‑end AI‑coding workflow, insisting that constraints must precede generation. He suggested three rule layers: (1) safety red‑lines that must be hard‑blocked, (2) business/architecture constraints that the model must obey, and (3) style preferences that can be flexible. He also advocated embedding requirement clarification, knowledge engineering and quality left‑shifting before any AI output.

When discussing organizational adoption, the panel noted that individual productivity gains (e.g., faster code writing) do not automatically translate to enterprise‑level efficiency. Liú proposed three metrics—average task time, pass‑rate of output verification, and tool reuse rate—to evaluate true impact. Xióng added system‑level metrics such as automated‑closure rate, experience‑capture rate and reduction of repetitive work.

The discussion concluded with a forward‑looking view: over the next three years AI agents will likely replace highly standardized, rule‑driven tasks (code comments, unit tests, simple APIs, documentation, routine ops checks), while high‑level design, safety, compliance and business value decisions will remain human‑led. The panel’s three takeaways were: (1) invest in AI‑native end‑to‑end development pipelines, (2) build long‑term memory and audit mechanisms for operational agents, and (3) define AI‑native workflow standards to gain a competitive edge.