In progress.
I’m trying Claude Opus 5.5 as the orchestrator and DeepSeek as the workers in T3 Code.
Dax
Opus hands implementation to DeepSeek subagents and reviews the diff once. Three lines from Dax’s write-up:
- we have long running sessions per project
- be strict about scope when creating a task
- ask for proof of work instead of a summary
Copy the proof-of-work prompt. Dax’s original post.
Drew Breunig
Drew uses cheaper models for most factory work and also runs Meta Spark in a modified Hermes agent. On the factory work and the pricing:
- the bulk of my tokens come from less complicated work in a SW factory (ds4-flash, glm 5.3)
- bills by concurrent lanes, not tokens