[bob1029]

> we’ve made GPT 5.6 Sol the default model powering every Ploy workspace

I would consider Luna for parts of the workload that touch actual tools. It is surprisingly capable and it runs fast.

Sol is great at talking to the human and orchestration of agent calls, but it's just too expensive to use everywhere.

You can get 5 Luna runs for the cost of 1 Sol run. Statistically speaking, going from one to five samples is a pretty big deal.

[Tadpole9181]

The problem I always run into with subagents is that they are isolated. This is a double-edged sword, as it keeps context down and lets them "focus", but it often means they must do their own research to continue to do work given to them, which eats uncached tokens.

So depending on how heavily agents are used on what tasks, it's entirely possible that you get worse work for more cost.

[bob1029]

> they are isolated

This is a feature if your goal is to obtain many samples. Independence is critical. This makes it easier to accurately model the uncertainty of a decision.

[taspeotis]

I feel Claude Code has added (and removed?) a feature that forks a subagent from the parent context, so it’s still isolated but it’s more of a continuation of what you were doing in one narrow direction and then it dies. Rather than a blank slate with a prompt of what to do.

[auspiv]

Statistically speaking if each part of the Luna run has a 90% chance of being correct, 5 of those is 0.9^5 = 0.59 = 59%. Or one Sol run being maybe 95% correct? Exact numbers vary of course. But then again having sol verify at end may be cheaper.

[bob1029]

The goal is not 100% correctness. The goal is to demonstrate the current amount of variance / uncertainty to the planning agent.

If the planner sees that 4/5 Luna runs resulted in approximately the same summary, it may conclude that variance is low and that it is over the target. If all Luna runs are different, the planner can conclude that additional research rounds are required.