Ambiguous success criteria. "Keep if results are better" gives no threshold. Write "keep if val_bpb decreases" instead.
Missing crash handling. Your agent needs to know: fix simple crashes (typos, missing imports) and move on from broken ideas.
Over-specifying implementation. "Insert a layer after attention" removes the agent's ability to find its own solution. Name the goal, not the steps.
Leaving VRAM unconstrained. Without any memory guidance, your agent will scale the model until it OOMs on every experiment.
Forgetting the NEVER STOP directive. Without it, your agent will politely ask if it should continue. At AM, nobody answers.