长程智能体可靠性：99%单步成功率为何不够
99% reliability sounds great but you should really do the math.
Larsen Cundric 指出，单步成功率 99% 时，10 步任务全部无失败完成的概率为 90%，100 步则降至 37%；若每次失败重试一次且失败相互独立，37% 可升至约 99%。但重试并不简单：响应超时可能导致动作重复执行，失败也可能改变系统状态，失败之间未必独立。因此长程智能体不只是 LLM 问题，围绕模型的系统同样关键。
正文
当前显示原文，中文全文译文尚未就绪。
99% reliability sounds great but you should really do the math.
If each action succeeds 99% of the time, a 10-step task has a 90% chance of finishing without a single failure. At 100 steps, that drops to 37%.
Now give each failed action one retry. Assuming independent failures, that 37% jumps to ~99%.
Thats a huge difference. But retries aren’t always that simple.
What if an action succeeded but the response timed out? Retrying could execute it twice. What if the failure left the system in a different state? They are probably also not independent.
You need to know what actually happened before deciding what to do next.
This is why long-horizon agents aren’t just an LLM problem. Better models are definitely needed, but the systems around them matter just as much!