This post, written in Chinese by Phodal Huang, starts from the claim that context awareness has always been central to AI-assisted programming and that, in 2025, with models no longer the bottleneck, obtaining the context a task actually needs will decide whether an AI assistant succeeds. Client-side assistants gather that context from the workspace and the editor or IDE; the post argues that an important trend for the year is for the server side — DevOps platforms and platform engineering — to become aware of the context tasks need as well. It summarises patterns for such “AI-friendly” platform engineering drawn, it says, from the author’s team’s work with clients and from their implementations in their AutoDev tool.
The author states that he is not a platform-engineering expert and that his understanding comes from AI coding and from implementing DevOps in enterprises. He divides the subject into five areas — platform user touchpoints, platform knowledge, a developer view built on standardised APIs, agent management, and AI metrics — and concludes that the core of an AI-friendly architecture is using platform engineering to supply AI coding assistants with rich, structured context through standardised tools, components and processes at every stage from requirements to deployment.
Key Points
- Before adding complex AI features, the post argues, enterprises should first invest in basic automation and standardisation — solid infrastructure as code, CI/CD and basic observability — and standardise core development and deployment flows through clearly defined “golden paths”.
- Most organisations have only latent assets rather than ready-made ones, and these must be digitised, standardised and componentised before an LLM can use them. One proposed pattern generates complete applications from natural language on top of golden-path templates; another supplies templated components to downstream AI assistants.
- For platform knowledge, the post proposes a platform’s built-in AI assistant as a first “trial field” (citing GitLab Duo), a “digital thread” linking requirements, design, code, tests and deployment records, and — as an experiment of the author’s own — a “context chain” built on a knowledge graph of domain concepts so that a developer can move from a requirement to its code, tests, deployments and documents and back.
- It describes pre-generated context, citing DeepWiki and Context7, as a way to avoid regenerating commonly needed context each time.
- It calls for semantically explicit, standardised “intelligent APIs” that treat AI agents as primary users, citing GitHub Copilot Extensions and the Model Context Protocol, which it calls a relatively mature standard despite its many problems. As a possible direction it suggests “AI Agent as Code” (AgaC), defining and versioning every element of an agent and its environment in code, by analogy with infrastructure as code.
- It suggests positioning the internal developer platform as the hub for integrating, configuring and governing AI assistants, and calls for observability specific to agents, such as detecting agents stuck in loops, misusing tools or drifting from their goal.
- On metrics, it argues that measuring speed alone can mislead because verifying AI suggestions takes time, and proposes breaking down time in the IDE, contextualising suggestion acceptance rates and measuring verification and cognitive overhead. It presents a metrics table that it says was designed by DeepResearch from DORA, SPACE and DevEx and which the AI named the “SPARE” framework.
- Until AI-specific security tooling matures, it recommends relying on classic static and dynamic analysis, code review and security requirements in prompts to mitigate vulnerabilities in generated code.
Context
The post is a catalogue of patterns and examples rather than an evaluation; several of its patterns are explicitly marked as experimental or potential directions, and it notes that the right form differs with an organisation’s size, technology stack and team culture. Its closing outlook expects agent-as-code practices and a protocol like Google’s A2A (Agent2Agent Protocol (A2A)) to become industry standards for connecting agents across systems.