Deciphering the new AI lingo: loops, seat belts, squads, hill climbing… oh my!

It might be overwhelming to see all the new vocabulary being introduced in software development these days thanks to the AI ​​tools that are introducing them… all the time.

Some of these new words describe useful patterns that people have recently been using, others are just fancy names on top of things that already exist, and some are still being actively defined as we speak.

In our latest episode of the GitHub Podcast, Marlene Mhangami, GPS, and I discussed some of the AI ​​terms that developers are learning right now: loop design, Ralph loops, squads, harness design, hill climbing, forward deployable engineers, closed models, open weights, and open source models.

If you’re a reader rather than a listener, here’s a guide to what these terms mean, why they matter, and how to think about them.

Listen to the full episode below! 👇

Designing Loops: Moving Beyond One-Time Prompts

Cyclic engineering is the practice of designing repeatable systems around agents, rather than manually asking them to perform one task at a time.

A simple example: Instead of asking an agent every morning to review new issues, summarize them, and suggest fixes, you create a cycle that runs on a schedule. This loop can identify issues, escalate them to an agent, review the output, and escalate anything that is stuck. This is a glorified cron job built into artificial intelligence.

Ralph Cycles: Cyclic Engineering’s Crude Cousin

The Ralph Loop is one implementation of this “loop” concept: you give an agent a detailed task, often from a document or product requirements specification, and make him work until the job is done.

This can be useful, especially for breaking large tasks into repeatable plan-do-check cycles. But on the other hand, it can also be expensive and inefficient since each iteration uses more tokens, more context, and more computation.

Loop design aims to make this pattern more structured so that you don’t have to keep asking the agent to “try again.” A well-designed loop adds primitives such as skills, observability, validation, routing, and checkpoints.

Squads, fleets and multi-agent workflows

While cycles define a workflow, “squads” and “fleets” describe how multiple agents can participate in that workflow.

A squad is a group of agents with different roles. They often reflect the real team. One agent can plan, another agent can test that plan, a third agent can implement it, a third can test it, and a third can verify it.

A fleet refers to parallel agents simultaneously working on tasks. You can have a squad working in the fleet in parallel, or sequentially.

This approach allows different agents to perform different parts of the process, and you can fine-tune and specialize each one with specific skills to improve efficiency.

The main idea is parallelization and specialization. Instead of one agent trying to do everything, different agents can perform different parts of the development process.

Straps: system around the model

Beyond what the model generates, the harness is everything that surrounds it, making it useful in your workflows.

These could be tools, permissions, memory, context, orchestration (etc.) that determine the behavior of the model. If it helps, remember: the harness is named after horse harnesses. Horses are like models that can run in the wild, and harnesses help guide the horse’s weight safely while performing tasks. Take it?

In any case, a good example of software is GitHub Copilot. It connects models to codebases, editors, pull requests, terminals, etc.

When you hear the term “wiring design,” it is the work of designing and improving the system surrounding the models.

Climbing the Hill: Improving Agents with Feedback

The term “hill climb” is used to describe the process of improving agents and harnesses over time.

This could mean, for example, using scores to measure whether the agent produces the correct result (and then adjusting the parameters until the results improve).

Or, as another example, if your agent has to review pull requests, Hill Climbing can check that it actually finds meaningful bugs and makes useful recommendations, and adjust the tools to improve this.

Advanced Engineer: A familiar role with a focus on artificial intelligence

The promising engineer job already exists, but AI branding makes it edgy and new. Now this is a customer facing software engineer, or a sales engineer, or a solutions engineer, often specializing in artificial intelligence.

If you haven’t seen these job titles before, this person typically works closely with customers to implement or adapt technical solutions to their environment. When we talk about artificial intelligence, this means helping teams integrate artificial intelligence tools, workflows, agents, etc. into existing systems.

Closed models, open weights and open source models.

Not all models are used equally.

Private models are accessed through an API or hosted product. Developers can use the model, but do not have access to the underlying weights, training data, or training process. The big, famous edge models you hear about are often closed models.

Open weight models make available model weights (which are like dials that determine how important certain inputs are). Developers can download and run these models, often locally or on their own infrastructure. But to be clear, the dataset and training method may not be fully available.

Open source models go even further: the model, code, data, and training process are available for inspection, reuse, and modification.

The more open the model, the more you can run it, customize it, test it, and trust it.

Conditions are constantly evolving

This is just a sample of some of the terms we hear often today. Some will remain, others will fade from our memories, and others will be replaced by better formulations as the industry evolves.

Don’t worry about falling behind on the buzzwords. These are just words, and what is more important is that there is practice behind them! Ask yourself whether workflows can be reliably repeated, how you validate tasks, how people should (or shouldn’t) intervene, how much you can rely on the model, and how you can improve your system. A new era of development has arrived and best practices still matter!

Posted by

Cassidy is the Senior Director of Developer Advocacy at GitHub. She enjoys building software, consulting for startups, and teaching developers how to build better. She has a weekly newsletter at cassidoo.co/newsletter where you can get her news, programming problems, and jokes in your inbox!