GitHub Standardizes AI Engineering Lexicon for Agentic Systems

GitHub has released a comprehensive guide and podcast episode to decode the emerging terminology used by developers building AI-native applications. The initiative seeks to standardize the language surrounding agentic workflows and system architecture, distinguishing genuine new engineering patterns from rebranded concepts.

According to the GitHub Blog, the industry is moving away from one-shot prompting toward "loop engineering," which involves designing repeatable systems around agents. These loops function as AI-native cron jobs, allowing systems to move beyond single-turn interactions. Within this framework, the guide identifies "Ralph loops" as a brute-force implementation where an agent works repeatedly on a task until completion, though GitHub notes this approach can be expensive in terms of compute and token usage.

The Architecture of Control

Central to this new lexicon is the "harness." A harness encompasses the entire system surrounding a model—including its memory, permissions, orchestration, and available tools—that renders the model useful within a specific workflow. Cassidy Williams describes the term as being aptly named after horse harnesses, noting that while models can "run wild," a harness directs that power safely to complete tasks.

For those scaling these systems, GitHub categorizes multi-agent workflows into two distinct structures: "squads" and "fleets." Squads consist of groups of agents assigned different specialized roles to collaborate on a goal, while fleets refer to parallel agents working on tasks simultaneously to increase throughput.

Iteration and Accessibility

To refine these systems, developers are employing "hill climbing," an iterative process of improving both the agents and their harnesses through continuous evaluations and feedback loops. This systematic approach to optimization marks a departure from the trial-and-error nature of early prompt engineering.

GitHub also clarifies the often-confused terminology regarding model openness. The guide tiers accessibility into three categories: Closed models, which are accessible only via API; Open weight models, where weights are available for local execution; and Open source models, which provide full access to the code, training data, and the training process itself.

Why the Shift Matters

This transition in language signals a broader industry shift from "prompt engineering" toward "AI engineering." The focus is migrating from the raw output of a model to the robustness of the surrounding system. By standardizing terms like harnesses and loops, developers can more effectively communicate complex architectures and build more reliable, repeatable AI software.

What's Next

As these agentic patterns mature, the industry will likely see a greater emphasis on the evaluation frameworks used in hill climbing. Developers should watch for how these standardized definitions influence the development of new orchestration tools and whether "Ralph loops" are replaced by more token-efficient iterative patterns.