← All writing

Routing work with a review step

I explored task-specific model choices while keeping bounded tickets, verification and independent review in the workflow.

By October 21, I was using Opus to coordinate work and looking for ways to make better use of my available subscriptions. I wanted smaller tasks assigned according to their difficulty, instead of every worker inheriting the coordinator’s model.

My working preference was Opus for coordination, with Sonnet retained for critical work and Haiku available for smaller tasks. I was also trying other models for implementation and review. Those were choices for my workflow at the time. I hadn’t run a controlled comparison that established a universal order of capability.

I was particularly interested in the model labelled Ox Alpha for code review. I’d liked its output and wanted to keep evaluating alternatives as they appeared. A preference like that can guide the next experiment, but it needs room to change when the results change.

The handoff requirements stayed fairly concrete. Before sending work to another model, I wanted a ticket with the goal, permitted files and acceptance commands. Critical changes needed appropriate review and a fallback if the chosen model couldn’t finish the job.

One small exercise in the thread showed why the review step belonged there. The assistant reported that an implementation of a slug conversion function passed its initial tests. An independent reviewer then found a word-boundary bug those tests hadn’t covered. A new regression check and another implementation attempt addressed it.

That’s a useful example of a missed case, with a limited scope. It doesn’t establish that the implementing model was weak or that the reviewer would always catch similar problems. It does show why passing the tests already written wasn’t enough for that particular change.

Review latency mattered too. One reviewer took long enough that a time limit became part of the proposed workflow. An extra opinion has a cost even when access to the model is free.

I also asked for short wrapup reflections on model use and token efficiency. I wanted later routing decisions to draw on what had worked in actual tasks. At the end of this setup work, the system was described as ready for a first real trial. Broader claims about savings or reliability would have to wait for that evidence.

Written by . Follow via RSS ↗