VRAM is a soft constraint in the default program.md. The directive reads: "Some increase is acceptable for meaningful val_bpb gains, but it should not blow up dramatically."
This is intentionally vague. A hard limit like "never exceed GB" prevents your agent from trying larger models that might be worth the memory cost. A soft constraint lets the agent make tradeoffs. It can use GB more VRAM if val_bpb drops by , but it shouldn't double memory usage for a gain.
Your agent logs peak_vram_mb for every experiment, so you can review the tradeoffs in results.tsv after the run.