O’Reilly Radar argues that AI coding systems need governance before generation begins, not just better prompts after the fact. Related documents from AGENTS.md and Anthropic show how project-level instruction files are becoming a practical way to give coding agents persistent context, build steps, conventions, and c...
O’Reilly Radar argues that governance for AI-assisted software development should begin before code generation, with project context, constraints, and threat models shaping what an agent is allowed to produce.
In “Context as Code,” O’Reilly Radar frames the problem as one of abundance: as AI systems make syntax cheaper to generate, the limiting factor becomes architectural control. The article’s central claim is that teams should not treat prompts as the main control surface for AI coding. Instead, it argues for build-time boundaries that make structurally invalid code less likely to enter a system in the first place.
That framing reflects a broader shift in AI coding practice. The Medium article “Prompt Engineering is Dead. Long Live Context-as-Code,” published by Artificial Intelligence in Plain English, describes Context-as-Code as replacing static, one-off prompts with version-controlled Markdown files such as CLAUDE.md, AGENTS.md, and SKILL.md. In that account, these files guide AI agents through onboarding, configuration, and runtime workflows.
The common idea is straightforward: if AI coding tools rely on context, then teams should manage that context with the same discipline they apply to source code, tests, and documentation.
The AGENTS.md project, published by the Agentic AI Foundation, presents AGENTS.md as an open Markdown format for giving coding agents project context. According to the AGENTS.md documentation, the file can include build steps, tests, and conventions, and is supported across tools including OpenAI Codex, Cursor, Google Jules, and other agent tools.
Anthropic’s Claude Code documentation describes a similar pattern for CLAUDE.md. Anthropic says CLAUDE.md is a source-controlled project instruction file that gives Claude Code persistent context across sessions. The documentation says teams can use it to record coding standards, workflows, and project architecture.
These examples point to a practical reason Context-as-Code is gaining attention: AI coding systems often need information that is too important to leave buried in chat history. Build commands, test requirements, architectural boundaries, naming conventions, dependency rules, and security expectations can be written down, versioned, reviewed, and reused.
O’Reilly Radar’s argument is not that instruction files magically solve AI reliability. Its point is more constrained: effective governance should start upstream, before an agent produces code. That means expressing intent, constraints, and threat models in the agent’s working context rather than trying to inspect every generated output after the fact.
For software teams, that distinction matters. A prompt can ask an AI system to “follow the architecture,” but a project context file can specify what that architecture is, how to run tests, which patterns are forbidden, and what conventions must be preserved. When those rules live in a repository, they become part of the project’s maintainable engineering surface.
The AGENTS.md and Anthropic documents also suggest why Markdown has become a common vehicle. It is readable by developers, easy to version, and simple for tools to consume. That does not make it a complete policy system, but it gives teams a shared place to encode operational knowledge that might otherwise remain informal.
The sources do not show that Context-as-Code eliminates hallucinations, security vulnerabilities, or flawed architectural decisions. They do show an emerging practice: teams are moving durable project knowledge out of ad hoc prompts and into source-controlled instruction files.
That practice may be especially useful where multiple AI coding tools are used on the same repository. AGENTS.md positions its format as open and cross-tool, while Anthropic documents CLAUDE.md specifically for Claude Code. The coexistence of these files suggests that teams may need to manage both tool-specific and tool-neutral context.
The near-term implication is modest but important. AI coding governance is becoming less about crafting the perfect chat message and more about maintaining the environment in which generation happens. If syntax is easy to produce, the harder engineering task is defining the boundaries that keep generated code aligned with a project’s architecture, tests, and standards.
O’Reilly Radar argues that governance for AI assisted software development should begin before code generation, with project context, constraints, and threat models shaping what an agent is allowed to produce.
From prompting to architectural control In “Context as Code,” O’Reilly Radar frames the problem as one of abundance: as AI systems make syntax cheaper to generate, the limiting factor becomes architectural control.
The article’s central claim is that teams should not treat prompts as the main control surface for AI coding.
Continue reading
