The second day of the AI Engineer World’s Fair (AIEWF) revolved around loops: repeated, agent-driven workflows and the automation patterns that arise from them. The opening keynote by swyx, titled “Loopcraft: The Art of Stacking Loops,” framed the discussion around layering loops and described AI engineering’s evolution since 2022 from chat to tools to goal-oriented automations.
Agents and looping patterns
Allie Howe of Keycard introduced the day’s main track, Software Factories, referencing Geoffrey Huntley’s idea that restarting an AI coding agent against the same specification can turn it into a persistent worker (the so-called “ralph loop”). Alexander Embiricos and Romain Huet from OpenAI focused on Codex and argued that chaining multiple agents through loops can boost productivity, especially when agents are connected not only to the task but to the purpose of the task and to follow-up steps like review and deployment.
Peter Steinberger (OpenClaw, now at OpenAI) also emphasized loop design: he uses loops to manage agents and sees the primary challenge today as choosing what to pay attention to. He suggested the future lies in better loop mechanisms to help solve that problem.
Software factories: automating the full lifecycle
Tereza Tížková from Factory defined a software factory as the entire loop — the full lifecycle of developing software with autonomy. That includes collecting signals, reacting to user feedback and logs, prioritizing what matters, and orchestrating the work. Zach Lloyd of Warp argued that software engineering will become factory engineering, with loops aimed at improving the system itself. Both speakers stressed that agents will be the builders: as Lloyd put it, “you’ll be building the thing that builds the product.”
At Warp’s expo booth, Lloyd explained that a factory mindset means picking repos, choosing lifecycle segments to automate, and deciding how and when to bring humans into loops. He acknowledged that the word “factory” can sound mechanistic to developers, but maintained that the discipline still requires problem-solving and that manual coding will be increasingly hard to justify as these systems accelerate development.
Forward Deployed Engineers (FDEs)
A prominent theme was the rise of Forward Deployed Engineers, sometimes called agent engineers. In an interview, Natalie Meurer, Head of Agent Engineering at Sierra, said much customer-specific work happens at the orchestration layer rather than in the base models themselves. Pauline Brunet, Cursor’s VP of Forward Deployed Engineering, positioned FDEs as partners who co-design and co-build AI software factories with organizations, transforming design, development and maintenance across the software lifecycle.
Open source AI and local models
Open source AI was another major trend. Zixuan Li of Z.ai spoke remotely about GLM-5.2, his company’s flagship model for long-horizon tasks, and introduced ZCode, a harness intended to support frontier models; he compared aspects of this work to OpenAI’s Codex. Thomas Wolf from HuggingFace interviewed Olive Song of MiniMax about their recently released open-weight model M3.
Ahmad Osman, founder of Osmantic, which builds open source tooling for deploying and operating local AI, noted that open models have improved significantly. He argued that more efficient architectures and incremental improvements compound, and once frontier labs prove a capability, the open-source community can often work backwards to reproduce it more efficiently.
Conclusion
Day two’s main trends were loops and their practical use, the software factory concept, the growing role of FDEs, and the rise of open-source and local AI models. Speakers suggested that combining agents with looped workflows could fundamentally change software development processes, while open models and local deployment tools are making AI more broadly accessible. Coverage will continue with analysis from day three, and the event’s keynotes are available to watch on YouTube.
