Found 1 document(s)
Page 1 of 1
-
When an LLM agent's move fails, the harness writes the move and the failure message into the transcript and asks again. In tool calling, this record makes small models more likely to repeat the failed action. We test whe...