Are we teaching coding agents to be productive, or just teaching them to spend?
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一位开发者注意到，当编码代理遇到困难时，它们通常会采取重复尝试和工具调用的策略，这导致了成本的增加，而这些成本往往被订阅或大额代币额度所掩盖。他们指出，代理会话中一些昂贵的部分，例如失败的编辑或重复读取文件，其成本可能超过实际任务或更改的成本。…
I've noticed something weird with coding agents lately.
When an agent gets stuck, the default solution seems to be:
more context -> another attempt ->more tool calls -> retry -> more context -> try again.
And because the result is usually hidden behind a subscription or a big token allowance, you don't really feel the cost while it's happening.
The funny part is that some of the most expensive parts of an agent session aren't necessarily the useful parts.
A failed edit followed by 3 correction attempts can cost more than the original task.
An agent repeatedly reading the same files can cost more than the actual change.
And sometimes giving the agent a better model is cheaper than forcing a cheaper model through five retries.
So I'm starting to think "tokens used" is a pretty bad metric for measuring whether an agent is being productive.
Maybe the better question is:
How much did the agent spend to produce work that I actually accepted?
Curious how other people think about this.
Do you track agent costs at all, or do you basically just use whatever allowance your plan gives you?