Decoding the new AI lingo: Loops, harnesses, squads, hill climbing… oh my!

Loop engineering involves creating repeatable systems for AI agents, rather than manually prompting them for each task. This concept helps manage larger tasks by executing them in cycles, often with validation and escalation steps. Ralph loops, a brute-force version of loop engineering, involve giving an agent a detailed task and having it complete it until finished.

However, this approach can be expensive and inefficient due to increased token use and compute requirements. Loop engineering aims to structure this process better by adding elements like skills, observability, validation, routing, and checkpoints. Squads and fleets represent multiple agents working together in a workflow. Squads are groups of agents with specific roles, while fleets refer to parallel agents working on tasks simultaneously.

Harnesses encompass the tools, permissions, memory, context, and orchestration surrounding AI models, directing their behavior. Harness engineering focuses on designing and improving this surrounding system. Hill climbing describes improving agents and harnesses over time through feedback, such as measuring output quality and adjusting harnesses accordingly.

A forward-deployed engineer is a customer-facing software or solutions engineer with an AI focus, helping teams integrate AI tools into their existing systems. Closed models, open weight models, and open source models differ in accessibility. Closed models are accessed via API or hosted products, while open weight models provide model weights for developers to run locally or in their infrastructure.

Open source models go further by making the model, code, data, and training process available for inspection, reuse, and modification. The terminology surrounding these AI concepts is still evolving, and it's essential to focus on understanding the underlying practices rather than getting caught up in buzzwords.