Why this matters

Ant Group will use AICon Shenzhen to explain how it is moving from AI-assisted coding to production-grade AI delivery, a shift that reflects a broader change in enterprise AI adoption.

The conference is scheduled for August 21–22 in Shenzhen, with the full agenda now published. Its tracks cover agent engineering, large-model infrastructure, AI-native development, embodied intelligence and other topics. The central question is no longer simply whether large models can generate useful code, but whether AI systems can operate reliably inside complex software delivery environments.

Liu Renquan, a senior technical expert at Ant Group, will speak in the track titled “AI-native new paradigm: Coding Agent reshapes the full software development process.” His talk, “Ant’s AI-driven production-grade software delivery infrastructure and practice,” focuses on the company’s exploration of end-to-end AI development and delivery. According to the event material, in many Ant teams, AI-generated code already accounts for more than 90%. That makes AI coding less of a moat and pushes the bottleneck toward AI Delivery: requirements handling, coding, evaluation, deployment, release control and production issue resolution.

From DevOps to agentic delivery

Liu is Ant Group’s CIO chief architect, head of the R&D efficiency Agentic Delivery team, and the lead architect for Ant Platform Business Group’s AI R&D infrastructure project. He has worked in R&D efficiency for nine years, participated in building Ant’s R&D efficiency team from the ground up, and led the architecture and implementation of a major internal one-stop DevOps platform for server-side development.

DevOps is a software engineering approach that connects development, testing and operations to improve the speed and stability of software delivery. In the AI era, the challenge for Ant is not just adding model calls to existing tools. The harder task is deciding which parts of more than a decade of engineering infrastructure should remain, which should be redesigned, and which workflows should be delegated to agents.

The talk will introduce Ant’s ADS architecture and lessons learned. One of its key design shifts is that agents are treated as first-class citizens. In this context, an agent is a software entity that can understand objectives, call tools and execute tasks. Ant’s stated principles under this view are weak platformization, CLI First and Code First. Weak platformization points to a more decentralized approach; CLI First emphasizes command-line interfaces as a standardized capability layer; Code First means turning engineering actions into machine-readable and executable artifacts.

The infrastructure stack

The disclosed ADS-related infrastructure covers several layers of AI-native delivery.

• Ama: infrastructure for AI-era requirements management and R&D workflows. It supports agent collaboration and AI workflows through a unified multi-agent runtime, allowing engineers to schedule multiple agents in parallel around the clock and coordinate long-running tasks across cloud and local environments.
• Poolab: a cloud-based development-agent environment that supports concurrent execution at the scale of tens of thousands of agents.
• ACLI: Ant’s unified For-Agent CLI standardization system. It is designed to address fragmented command-line tools, inconsistent standards, weak security capabilities, and disconnected build, release and distribution chains.
• ADE: an environment layer that makes evaluation and pre-release environments creatable, verifiable, fresh and routable for agents. It aims to turn fragmented capabilities such as environment centers, Yuntu, GYP and MOSN into an orchestrated loop, reducing evaluation traffic failures, environment drift and manual troubleshooting.
• Agentic CI/CD: an AI-driven approach to build, deployment, traffic switching and gate control. It provides a unified intent-based delivery entry point, hides differences among deployment platforms, and supports pipeline monitoring, root-cause diagnosis and traffic switching.

CI/CD stands for continuous integration and continuous delivery. It is the automated process that builds, tests and moves code changes toward production. Ant’s agentic version focuses on translating user intent into executable delivery steps while reducing manual work across heterogeneous deployment systems.

Production practice and unresolved tensions

The main production example is the “Afu 7×24 Harness” practice. A harness here can be understood as an end-to-end loop that connects task execution, validation and feedback. In Afu’s BadCase repair scenario, ADS completed a production end-to-end harness spanning attribution, solution design, coding, evaluation and release.

The disclosed results are notable:

• More than 300 BadCases have been resolved.
• Efficiency improved by more than 3 times, with the cycle reduced from 12.6 days to 3 days.
• AI-led work accounted for over 95% of the full process.
• Humans mainly made decisions and judgments at key checkpoints.
• Beyond Afu, the practice has covered more than 10 business teams, including Lingguang and TeamClaw.

The material also points to practical pain points. Ant has accumulated extensive development infrastructure over more than a decade, and not all of it maps cleanly to the AI era. The company had to make trade-offs over what to retain and what to discard. Another tension comes from the falling cost of AI coding: business teams may build their own tools, creating a conflict between decentralized innovation and centralized infrastructure products.

What comes next

Ant’s case suggests that the next phase of AI software engineering will be defined less by code generation alone and more by production system design. Enterprises need agents that can work with permissions, environments, evaluations, releases, monitoring and rollback mechanisms, not isolated assistants that only produce code snippets.

The near-term direction for AI-native development is likely to center on three capabilities: governable agent platforms, standardized AI-facing environments and CLI ecosystems, and end-to-end production feedback loops. Companies that turn AI from a local productivity tool into stable delivery infrastructure will be better positioned for the next shift in software engineering.