AI Development Terminology Explained: Loop Engineering, Squads, Harnesses, and Other Recent Concepts

With the proliferation of AI tools, new terminology is rapidly emerging in the software development field. A GitHub podcast episode summarized the meanings and importance of key AI terms that developers are currently learning, including loop engineering, Ralph loop, squad, and harness engineering.

Loop Engineering and Ralph Loop

Loop engineering is the practice of designing repeatable systems instead of manually prompting agents. For example, scheduling a loop to review issues and propose fixes every morning automates the task, making it function like an AI-native cron job.

The Ralph loop is one implementation of this loop concept, where detailed tasks are assigned to an agent and repeated until completion. It is useful for breaking down large tasks into plan-execute-verify cycles, but it can be inefficient due to the consumption of tokens and computing resources with each iteration. Loop engineering aims to reduce this inefficiency by adding structured primitives such as skills, observability, verification, and routing.

Multi-Agent Workflows: Squads and Fleets

A squad is a group of agents with different roles that share functions like planning, reviewing, implementing, testing, and reviewing, similar to a real team. A fleet refers to parallel agents performing tasks simultaneously. This allows for parallelization and specialization by fine-tuning each agent for specific skills, rather than having a single agent handle everything.

Harnesses and Hill Climbing

A harness refers to the entire system surrounding a model to make it useful in a workflow. It includes tools, permissions, memory, context, and orchestration, with GitHub Copilot being a prime example of a software harness. Harness engineering is the work of designing and improving this system.

Hill climbing describes the process of measuring agent output quality through evaluations and adjusting the harness to progressively improve performance. For example, verifying whether a PR review agent finds meaningful bugs and adjusting the tools accordingly.

Other Terms

A forward deployed engineer refers to a customer-facing engineer role that works closely with customers to integrate AI tools and workflows into existing systems. Additionally, models are categorized based on their release method: closed models accessible only via API, open weights models that release their weights, and fully open source models.