AI generated 90% of the code. Delivery cycles barely moved

One number tells most of this story, and it is not the one the team is proudest of. At the Yunqi Conference last September, Wang Shuxin of Alibaba's AMAP Large Model Application Platform reported a 53% AI code-generation rate across the technical-design and development phases. Half a year later, the same team reaches 80 to 90% and beyond. The number nearly doubled, while delivery cycles did not shorten and developer workloads did not decrease. "We discovered a puzzling fact: the efficiency gains were not obvious," Wang said in a talk published on the Alibaba Cloud Community blog.

The gap between those two facts is where the lesson sits. AI wrote more of the code and the software shipped at the same pace. Alibaba's own biggest model coded alone for 16 days, so raw generation ability is not what's missing. The diagnosis Wang's team landed on doubles as an argument against the metric most of the AI coding industry reports as progress.

Why a doubled metric bought nothing

The first reason is arithmetic the software industry has known since The Mythical Man-Month. Development is one stage in a long chain: product proposal, product-and-engineering review, solution design, development, code review, testing, integration, launch. Every stage carries communication costs, waiting time, and the chance of error. If you optimize coding by 50%, and coding is 30% of the chain, the overall gain is 15%.

The second reason is hidden cost. AI-generated code brings more code-review time, more debugging, more rework. The team's concrete case: AI changed the parameter order of a core interface, every unit test passed, and after launch it broke three downstream services. Tracking it down took a full day. Alibaba Cloud has also built tooling to shorten that kind of hunt; its STAROps plugin turns a one-sentence question in Qoder into a root-cause diagnosis.

The third is context. Large tasks do not fit in a model's head at once. A refactoring job touching a dozen front-end and back-end modules cannot survive a single conversation; the AI's attention gets scattered, and constraints stated early vanish. How far an agent can stretch across a big codebase is being tested in production, including the 600-file refactor moonshot.ai ran with Kimi Code CLI. Wang's conclusion is blunt: in legacy applications, AI programming has to move from "vibes" to "conventions," with clear acceptance criteria.

The fix: specification as source of truth

Wang defines vibe coding as "programming by vibes": casually tossing AI a few prompts and letting it generate thousands of lines of code in seconds. Fine for new projects and small scripts, he argues. Legacy applications are where it breaks: they carry historical baggage, implicit dependencies, and business knowledge embedded in the code, and a plausible-looking AI solution can be incompatible with all of it.

AMAP's answer, built on its Qoder tool, is Specification-Driven Development. The spec stops being a prose guidebook that drifts out of date. It becomes structured "intent code" that agents execute precisely, and the single source of truth the code must match. Qoder's Quest Spec mode walks developers through the details, and the acceptance criteria are where this gets testable. "The user is redirected to the home page after a successful login" is vague, Wang says. "After a successful login, the user is redirected to the home page within 3 seconds, and the home page displays the user's nickname" is something a test can actually check. The workflow runs in four stages:

Alongside SDD, AMAP applies what it calls Harness Engineering. The image is a wild horse: the model has enormous power, and without a harness you cannot mount it. The harness has four pillars. Context engineering keeps a single source of truth. Architectural constraints block violations before submission, so UI-layer code cannot touch the database layer directly. Qoder applies the same enforcement idea on the security side, with checks running from keystroke screening to cross-file data-flow analysis. Feedback loops: tests run, the agent reads error logs and self-corrects, and human bug fixes harden into rules. The last pillar is human oversight for what Wang calls the 5% of ambiguous logic AI cannot judge. Deployment closes the loop through MCP tools provided by Aone, Alibaba's internal CI/CD platform.

The metric problem no one is fixing

The lesson is not that AI coding failed. Code-generation rate was never the right dashboard. A team that nearly doubled its rate and gained nothing is proof that the number measures activity, not throughput. What AMAP tracks now is the whole chain, from requirement to launch, with AI as infrastructure rather than a glorified autocomplete.

Wang's open questions are honest about the distance left: spec generation still needs human intervention and should get more conversational; agent teams need richer collaboration, including multi-round iteration and dynamic role assignment; knowledge management needs smarter extraction and reuse. Until those improve, the gap between what generation rates promise and what delivery cycles deliver will keep showing up in PMO data.

Nobody ships a code-generation rate.