This is the real transformation path that a mature ICT enterprise carved out over three years.
I'm Zhu Jun, Chief Architect at AsiaInfo Technologies.
AsiaInfo has been deeply rooted in the telecom industry for over three decades, with nearly 10,000 R&D engineers. At this scale, "company-wide AI Coding" sounds more like a slogan than a goal — but this year, we actually pulled it off. From frontline engineers to architects, almost everyone is now collaborating with Agents in their daily work.
What I want to share isn't a polished slide deck. It's the real-world AI transformation path that a mature ICT enterprise navigated over three years.
Telecom carriers' BSS (Business Support System) and OSS (Operations Support System) projects make up a significant share of AsiaInfo's business portfolio.
These IT systems share two defining traits:
● Extreme system complexity: from 2G to 5G, legacy services never get retired while new requirements keep flooding in — every change ripples through everything.
● Extreme knowledge accumulation: thirty years of engineering know-how, scattered across senior experts' minds, hundreds of out-of-sync documents, and millions of lines of legacy code.
A new hire used to spend two to three months just getting familiar with a single module.
Meanwhile, telecom carriers are pursuing business transformation — strengthening the core business while building out a second growth curve. That puts a hard "efficiency growth" demand on AsiaInfo as their service partner.
We had to use AI to reorganize the production relationships.
After ChatGPT launched at the end of 2022, my team and I had a serious conversation about one question: what will future application software look like?
Our answer was: Cloud Native will inevitably evolve into AI Native.
The shift plays out across three layers:
● Infrastructure shifts: from container-centric PaaS to large-model-centric MaaS.
● Development paradigm shifts: from DevOps to AI Engineering — from humans writing code to humans steering Agents that write code.
● Application paradigm shifts: from SaaS to Agents, supported by three pillars — "business ontology knowledge + digital twin + intelligent agents."
Once the direction was clear, we baked "AI First" into every detail of how we work.
We started with "foundation model + RAG," feeding our own code and documents into the system. Then we brought in Qoder CN (formerly Tongyi Lingma) and Qoder. The capability boundary expanded from "code completion" step by step into requirements, design, testing, and bug fixing — gradually forming a complete R&D team.
This journey crystallized into three iron rules I now use for tool selection:
● Look at sustained investment in the foundation model: is the company really pouring resources into software engineering, and is it iterating faster over time? This determines whether you're buying a car that can keep upgrading.
● Look at the surrounding ecosystem: it's not just the model — it has to fit your organization, processes, and standards, and be flexibly configurable.
● Look at the security foundation: a trustworthy model and a trustworthy runtime are the entry ticket for clients like telecom carriers.
The hardest part has never been writing code — it's preserving knowledge.
Senior experts' experience and decades-old code logic used to be passed down through one-on-one mentoring. That was inefficient, and it's becoming unsustainable. Once we brought Agents in, the problem became even more direct: silicon-based and carbon-based employees have to collaborate, which means the Agents also need to "understand" our systems.
Our approach is to truly structure the business ontology knowledge so it can be stored in a form that both humans and Agents can read and reach consensus on.
Qoder's knowledge engine made a real difference here. The key is not "yet another RAG," but:
● AI-driven extraction and maintenance: instead of relying on senior engineers to manually build the library, the engine automatically extracts standards, API assets, and architectural traits from the codebase, and turns daily conversations into knowledge cards.
● Knowledge-friendly format: long documents are split into small cards and paired with hybrid retrieval, so Agents can hit the right note as quickly as flipping through index cards.
● Repository-level sharing: when a senior expert deposits knowledge into a repo, the entire team can share it.
In one sentence: senior experts deposit knowledge, while new hires and Agents share and reuse it together.
There are no "small projects" in carrier systems. Take one carrier's healthcare project: front office, back office, billing, and customer service are all wired together end-to-end, spanning 31 provinces plus the group HQ.
In the past, this kind of system ran on a serial R&D rhythm — touch one piece and you have to redo everything. The way we work now:
● Architects first define the overall boundaries, module breakdown, and interface contracts.
● Spec is used to write down constraints and roles — for example, "this Agent only handles localization, not code changes."
● Multiple Agents move forward in parallel, then everything is reassembled at the architecture layer.
We've polished this workflow most thoroughly in the deterministic scenario of bug fixing. Early on, we just dumped logs and user feedback at the AI, and accuracy was poor. Then we layered in constraints, expert experience, code context, architectural knowledge, role boundaries — and the success rate steadily climbed.
Human control over AI doesn't come from watching it. It comes from layering constraints and knowledge into the workflow.
The Qoder team's own sharing left a deep impression on me.
Their development cadence: in nine months, they shipped 60+ external releases and over 130 internal versions — averaging a new build every one to two days.
Why so fast? Because Qoder is built using Qoder:
● Product managers sketch prototypes, and design flows directly into code.
● New features are developed in Spec mode.
● Once a bug is filed, the environment and logs are automatically packaged and handed to an Agent, which reproduces, fixes, and verifies it on its own.
● Integration tests and regression tests — anything that can be handed to an Agent is handed over.
What humans keep is decision-making, steering, and continuous improvement.
This gave me an important insight — AsiaInfo is now building an "Internet of Agents": my Agent buys for me, your Agent sells for you, and behind it all sits a new layer of billing, protocols, and trust. It's an entirely new application paradigm, and only this kind of "self-developing-self" iteration speed can bring it to life.
We're shaping the corresponding product forms too. Qoder Wake, recently launched by the Qoder team, runs Agents on individual machines so that the bug-localization, bug-fixing, and final-testing Agents share context and coordinate with each other — that's the early form of a digital employee.
When you talk about delivery, you can't avoid the numbers. We've always measured R&D the way we run software engineering projects. After bringing in AI, all that changed was that we added one more dimension — "the contribution of digital employees."
● Coverage: BSS systems hit the highest coverage. OSS is slightly lower due to stronger deterministic constraints. Overall, around 80%.
● Activity: take the team serving one major carrier — 1,000+ registered engineers, with monthly active rate consistently above 90%.
● Code-level efficiency: per-engineer code output up by roughly 14.6%.
● End-to-end efficiency: across the full pipeline from requirements to testing, average uplift of 8.6%.
These aren't headline-grabbing numbers, but they're real. And every percentage point frees up significant headcount to work on innovation — that's the real value of bringing in AI.
If your enterprise is considering bringing AI in, here are three pieces of advice:
● First, model selection sets the ceiling: the foundation model is the bedrock. Without sustained investment, it won't take you far.
● Second, manage Agents like new hires: you need to give them a role, set up the environment, define the boundaries, and run evaluations — not just "install a tool."
● Third, drive the flywheel with data: without data, you can't tell where to adapt your workflow when the next generation of models arrives.
There's a quote from Steve Jobs that has stayed with me — Technology Alone Is Not Enough.
AI Coding gives us an opening: to free people from repetitive labor, so they can focus on work with more taste and more creativity.
From here on, your collaborator isn't just a carbon-based colleague — it's also a group of digital partners evolving alongside you.
Together, we achieve more.