AI and LLMs · Senior · Question 66 of 75
Why are long-horizon agents harder to evaluate than single calls?
- 1Errors compound over steps and depend on the environment
- 2They always finish faster than single model calls
- 3Their outputs are always shorter and simpler
Show the answer
Correct answer: Errors compound over steps and depend on the environment
Explanation
Evaluations need realistic sandboxes, checkpoints and budgets, and failures must be analysed step by step.