You can steer your agent toward specific experiments by adding hypotheses directly to program.md.
Weak instruction: "Try different optimizers."
Strong instruction: "MuonAdamW shows promise in prior work. Test variants of the lr scheduling component while keeping the Muon step logic consistent."
The difference: the weak version gives your agent no starting point. The strong version points at a specific area and says "dig here." Your agent still decides what to try, but it starts with a focused hypothesis instead of random exploration.
The best time to add directed hypotheses is after your first overnight run. You've seen what worked. Now guide the next run toward variations of those successes.