New AI Terms and Practices in Development
How do concepts like loop engineering, squads, and harnesses work in the AI world?
1What is Loop Engineering?
Loop engineering is the practice of designing repeatable systems for AI agents instead of one-off commands. For example, rather than manually reviewing new issues every morning, a loop can automate this process. The loop fetches issues, passes them to an agent, validates the output, and escalates stuck items. It functions like an AI-native cron job but can manage more complex workflows.
2Ralph Loops: The Brute-Force Approach
Ralph loops are an implementation where an agent takes a detailed task from a product requirements document and keeps working until completion. Useful for breaking large tasks into iterative plan-act-check cycles, but costly as each iteration consumes more tokens, context, and compute. Loop engineering aims to structure this process for better efficiency.
3Squads and Fleet Systems
Squads and fleets define how multiple agents can participate in a workflow. A squad consists of agents with different roles, often mirroring a real-world team. A fleet refers to parallel agents working simultaneously. This approach allows different agents to handle distinct parts of a process, each fine-tuned for specialization and efficiency.
4Harnesses: The System Around the Model
Harnesses include tools, permissions, memory, and context that make a model useful in workflows. GitHub Copilot is an example, connecting models to codebases, editors, and pull requests. Harness engineering involves designing and improving these surrounding systems to enhance model performance and integration.
5Hill Climbing: Improvement via Feedback
Hill climbing describes the process of refining agents and harnesses over time using feedback. For instance, if an agent reviews pull requests, evaluations check whether it finds meaningful bugs and offers useful recommendations. Tools are adjusted until results improve, ensuring continuous enhancement of AI systems.
6Forward-Deployed Engineers
Forward-deployed engineers work closely with customers to integrate technical solutions into their environments. The AI-focused variant helps teams adopt AI tools and workflows. This role resembles sales or solutions engineering but is rebranded with an AI emphasis, reflecting the evolving needs of tech integration.
7Closed vs. Open Models
Closed models are accessed via APIs, with no access to underlying weights or training data. Open-weight models release weights for local deployment but may withhold datasets or training methods. Open-source models provide full transparency, allowing inspection, modification, and trust-building through complete access to code, data, and training processes.
8Why These Terms Matter
These new terms reflect the rapid evolution of AI and software development practices. Concepts like loop engineering and multi-agent systems enable developers to build more efficient, scalable solutions. While the buzzwords may fade, the underlying practices will shape the future of the industry, driving innovation and productivity.
Original article: The GitHub Blog — Yapay Zeka Geliştirmede Yeni Terimler ve Uygulamalar