Is Claude Fable 5.1 spawning 126 agents for simple tasks for anyone else? It burned ~8.4M tokens on a 5-file i18n audit!
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。
一位订阅了每月100美元Max 5x套餐的用户报告称，Claude Fable 5.1在执行一个简单的5文件国际化审计任务时，意外地生成了126个代理，消耗了大约8.4M的tokens。该用户正在询问其他用户是否也遇到了类似的、未经明确请求的高并行代理生成问题，特别是出现了126这个确切的数字，这让他们怀疑系统中可能存在一个编排错误。
I’m on the $100/month Max 5x plan, and I’ve just had one of the strangest Claude Code experiences I’ve seen so far.
I asked it to audit the login flow of my project for i18n/localization issues.
So roughly 8.4 million tokens were consumed across two executions of a relatively simple localization audit.
This burned through multiple 5-hour usage windows and roughly half of my weekly Max allowance.
What concerns me more is that this doesn’t look like ordinary “agentic workflows use more tokens.”
The same workflow independently hitting exactly 126 agents twice looks much more like some kind of deterministic fan-out/orchestration problem.
For a while, Anthropic’s Fin AI support agent repeatedly told me there was no separate path to human Product Support, even after I explicitly asked for human review of the subscription impact.
Eventually, after opening another support conversation and framing this specifically as a reproducible product defect affecting a paid subscription, Fin finally put me into the human support queue.
I’m not expecting agents to be cheap, and I understand that parallel workflows can burn tokens quickly.
But 126 agents for a 5–6 file localization audit, twice in a row, feels completely broken.
Has anyone else seen workflows suddenly create 50, 100, or 126+ agents without explicitly asking for that level of parallelism?
Especially interested in whether 126 is showing up for anyone else, because seeing the exact same number twice is what really makes me think there’s an orchestration bug here.