Eighteen months ago every week brought a new way to chain model calls together. Most of those projects are now archived. What remains is a small set of frameworks with real maintainers, stable interfaces and, crucially, tracing that works.
What survived and why
The winners share a pattern: they treat the model as one component among many and keep the control flow in ordinary code. Libraries that hid everything behind a magic loop struggled the moment a team needed to debug a production failure at 3 a.m.
Key takeaways
- Three frameworks now cover most production deployments
- Observability became the deciding feature
- Teams are standardising on plain functions over prompt DSLs
This story is part of our ongoing coverage of ai. We will update it as more details emerge.
