Why Faster AI Coding Has Not Made Enterprise Delivery Faster

Introduction

Large language models can now produce code at impressive speed. That does not mean a product requirement can move from discussion to production at the same pace. Enterprise delivery still depends on design systems, internal knowledge, repository conventions, security checks, testing and coordination between different roles. In a presentation about Xiaohongshu’s AI Coding platform Muse, the company’s architecture team argued that coding is only one segment of the development process. Time saved there can be lost during review, rework and handoffs.

The real problem is end-to-end delivery

The material identifies three recurring obstacles:

• The model lacks enterprise context. It may not know internal components, business rules, repository structure or design standards.
• Relevant information is fragmented. Conversations, user memory, designs and code often live in separate systems, preventing a unified understanding of the task.
• Tools do not automatically form a reliable process. A collection of Skills and tools can still produce conflicting actions, lost context or tasks that cannot be resumed.

Muse is designed as more than a generator of disposable HTML prototypes. It aims to place product ideation, design exploration and production-oriented implementation on one context chain. Product managers, designers and developers can work around editable artifacts, compare alternatives and pass the resulting context to a Dev Agent that continues into the real repository.

Orchestration needs a control plane

Muse distinguishes work before a requirement becomes an engineering task from work inside the production codebase. The first stage gathers business and design context; the second must handle cross-repository understanding, coding conventions, validation and delivery constraints. To balance precision and generalization, the platform does not treat Agent Teams as a universal replacement for structured workflows. Workflows and Pipelines remain useful where ordering and behavior must be predictable, while Agent Teams are introduced when tasks can be decomposed, separated and dynamically coordinated.

The Harness layer provides the control surface around an Agent. It can inspect context before execution, monitor each turn, validate tool inputs and outputs, and pause before actions with side effects for human approval. Recoverability also requires more than preserving a transcript. A runtime should store structured task goals, plan versions, completed steps, tool evidence, approvals, resource versions and failure information. This makes it possible to pause, audit, resume and, where appropriate, roll back a task.

Why this matters

Muse points to a broader shift in AI Coding. The central question is no longer only which model writes code fastest, but which system can turn model output into dependable production work. The model sets the upper bound of capability; the Agent runtime, enterprise knowledge, validators and permission boundaries determine whether that capability is safe and repeatable in practice.

Humans are not simply removed from the process. Their work moves toward judging goals, comparing alternatives, supervising risk and applying product taste. Agents can expand the solution space and perform more actions, but only a closed loop of context, tools, state, verification and approval can turn Vibe Coding from a rapid prototyping experience into a credible engineering capability.

Source: InfoQ 中文