AI tools are introducing new vocabulary into software development. Some terms describe emerging engineering practices, while others rename established roles or concepts that are still evolving. This guide explains several expressions discussed by Cassidy Williams, Marlene Mhangami, and GPS on the GitHub Podcast.
Loop engineering and Ralph loops
Loop engineering means building repeatable systems around agents rather than prompting them manually for individual tasks. For example, a scheduled workflow could retrieve new issues, ask an agent to summarize and propose fixes, validate the result, and escalate unresolved cases. In practice, this is an AI-oriented version of a cron job.
A Ralph loop is a more direct implementation of the same idea. An agent receives a detailed task, often based on a product requirements document or specification, and continues through repeated plan, action, and checking cycles until the work is complete. This approach can help with large tasks, but repeated iterations consume additional tokens, context, and compute. More structured loops can add skills, observability, validation, routing, and checkpoints instead of simply asking the agent to try again.
Squads, fleets, and multi-agent workflows
A squad is a collection of agents assigned different roles, such as planning, evaluating a plan, implementing changes, testing, and reviewing. A fleet describes agents working on tasks in parallel. Squads and fleets can operate either sequentially or concurrently.
This arrangement emphasizes specialization and parallelization. Rather than asking one agent to manage every stage of development, teams can assign distinct responsibilities and tailor each agent with relevant skills.
Harnesses and harness engineering
A harness is the system surrounding a model that makes it useful in a workflow. It can include tools, permissions, memory, context, and orchestration. The term compares this surrounding system to a horse harness, which directs and controls the horse.
GitHub Copilot is one example of a software harness because it connects models with codebases, editors, pull requests, terminals, and other development tools. Harness engineering refers to designing and improving these surrounding systems.
Hill climbing and feedback
Hill climbing describes the ongoing improvement of agents and harnesses through feedback. Teams can use evaluations to measure whether an agent produces the desired output, then adjust the harness based on the results. For a pull request review agent, this might involve checking whether it identifies meaningful bugs and provides useful recommendations before refining its tools.
Forward-deployed engineers
Forward-deployed engineering is an established customer-facing role that has gained new attention through AI terminology. These engineers may also be called sales engineers or solutions engineers. They work with customers to implement or adapt technical systems, and an AI-focused version helps teams integrate AI tools, workflows, and agents into existing environments.
Closed models, open weights, and open source models
Closed models are provided through an API or hosted product. Users can access the model but not its underlying weights, training data, or training process. Many prominent frontier models are closed.
Open weight models publish the model weights, the parameters that determine the importance of inputs. Developers can often download and run them locally or on their own infrastructure, although the data and training methods may remain unavailable.
Open source models provide broader access to the model, code, data, and training process for inspection, reuse, and modification. Greater openness can increase opportunities to run, customize, and audit a model.
Why the terminology may change
AI engineering language is still developing. Some of these terms may become standard, while others may disappear or be replaced. The underlying practices matter more than memorizing buzzwords: reliable repeatability, task validation, appropriate human involvement, model limitations, and continuous system improvement remain central to effective engineering.