The move from agent loops to structured graphs, with the research behind it

I run an AI consultancy and wrote this up after a run of client work. Sharing because it might be useful, not just to drop a link. The shift I keep seeing in production agent work: from running agents as loops (point it at a task, let it repeat until done, like the Ralph Wiggum technique) to running them as graphs: named steps, defined edges, explicit state you can inspect and checkpoint. A few things stood out digging into this: - Durable execution engines (Temporal, Restate) mean a crash mid-run resumes from where it stopped, instead of restarting or repeating a step that already had a side effect. - A 2026 survey (arXiv 2603.22386) found that fully generating a workflow graph at runtime is usually overkill. The pragmatic sweet spot is one well-validated graph with a router picking the right path per task. - AFlow, which searches the graph-structure space with Monte Carlo Tree Search, beat manually designed workflows by 5.7% and other automated methods by 19.5%, and got a cheaper model to GPT-4o-level results at 4.55% of the cost. Full writeup with sources in the first comment (keeping this post itself link-free per the sub's rules). Curious whether others here are seeing the same shift, or still finding loops good enough for what they're building.