5D Library › 一人公司自动化架构、数字人直播漏斗与任务级人工智能套利 · English
第二章 智能体链条与创始人的排队
误差沿链条累积
所谓「硅基团队」通常是一条链:一个智能体做调研,一个起草,一个排版,一个发送。若每一步正确的概率是 ,误差相互独立,整件工作正确的概率是 。检查可以改变这一点,但要付出代价。下面设一次检查以 0.9 的概率抓出错误结果,成本是五分之一步。
推导
# Chapter 2. p03: Reliability of a chain of AI agent steps. A job passes
# through k steps; each step is right with probability p and errors are
# independent. Without checks the job is right with probability p**k.
# A check after a step catches a wrong result with probability 0.9 and
# sends the step back to be redone (costing one more step); a check costs
# 0.2 of a step. Compared: no checks, one check at the end (catching any
# error with probability 0.9, then the whole job is redone at most once),
# and a check after every step. Exact expectations, no simulation.
def no_check(p, k):
return p ** k, float(k)
def end_check(p, k, d=0.9):
ok = p ** k
caught = (1 - ok) * d
right = ok + caught * ok # one full redo after a catch
cost = k + 0.2 + caught * (k + 0.2)
return right, cost
def step_checks(p, k, d=0.9):
# per step: geometric redo until right or an error slips through
q = 1 - (1 - p) * d # step ends (right or slipped)
right_step = p / q
tries = 1 / q
return right_step ** k, k * tries * 1.2
if __name__ == "__main__":
print(" p k no check end check every step")
for p in (0.95, 0.98, 0.99):
for k in (5, 10, 20):
a, b, c = no_check(p, k), end_check(p, k), step_checks(p, k)
print(" {:.2f} {:2d} {:.3f} ({:4.1f}) {:.3f} ({:4.1f})"
" {:.3f} ({:4.1f})".format(p, k, a[0], a[1], b[0],
b[1], c[0], c[1]))
print(" (share of jobs delivered right; expected steps of work,")
print(" checks included)")
# p k no check end check every step
# 0.95 5 0.774 ( 5.0) 0.931 ( 6.3) 0.974 ( 6.3)
# 0.95 10 0.599 (10.0) 0.815 (13.9) 0.949 (12.6)
# 0.95 20 0.358 (20.0) 0.565 (31.9) 0.900 (25.1)
# 0.98 5 0.904 ( 5.0) 0.982 ( 5.6) 0.990 ( 6.1)
# 0.98 10 0.817 (10.0) 0.952 (11.9) 0.980 (12.2)
# 0.98 20 0.668 (20.0) 0.867 (26.2) 0.960 (24.4)
# 0.99 5 0.951 ( 5.0) 0.993 ( 5.4) 0.995 ( 6.1)
# 0.99 10 0.904 (10.0) 0.982 (11.1) 0.990 (12.1)
# 0.99 20 0.818 (20.0) 0.952 (23.5) 0.980 (24.2)
# (share of jobs delivered right; expected steps of work,
# checks included)每步正确率 0.95 时,10 步链条不加检查,交付正确的比例只有 0.599,20 步链条只有 0.358。只在末尾检查一次、抓到就整体重做,可把 10 步链条提高到 0.815,期望工作量 13.9 步;每步之后都检查,提高到 0.949,工作量 12.6 步,反而更省,因为错误在后续步骤叠加之前就被抓住了。即使每步正确率 0.99,20 步不检查也只交付 0.818。实用规则是:链条要短,在纠错便宜的地方检查,不要让长而无检查的流水线直接面对客户。
创始人是最后一个服务台
智能体解决不了的事会上交给创始人。上交随机到达,创始人一次处理一件,所以创始人是一个单服务台排队系统 (Kleinrock 1975)。设创始人每小时处理 6 件,每天工作 6 小时,上交比例为 。
推导
# Chapter 2. p04: The founder as the last server. AI agents handle the
# customer conversations; a share e of them is escalated to the founder,
# who answers 6 an hour (10 minutes each) during 6 working hours a day.
# Escalations arrive at random, so the founder is an M/M/1 queue with
# arrival rate lam = e * N / 6 per hour: utilisation rho = lam / 6, mean
# time from escalation to answer W = 1 / (6 - lam) hours, and the chance
# of waiting over 1 hour before the founder starts is rho * exp(-(6 -
# lam)). The queue is stable only while N stays below 36 / e a day.
import math
MU, HOURS = 6.0, 6.0
def queue(n, e):
lam = e * n / HOURS
rho = lam / MU
if rho >= 1:
return rho, None, None
return rho, 60 / (MU - lam), rho * math.exp(-(MU - lam))
if __name__ == "__main__":
for e in (0.02, 0.01):
print(" escalation share {:.2f}: stable below {:,.0f} a day"
.format(e, MU * HOURS / e))
for n in (600, 1200, 1500, 1700, 1750, 3000):
rho, w, p = queue(n, e)
if w is None:
print(" {:5,d} a day: rho {:.2f}, queue grows without"
" limit".format(n, rho))
else:
print(" {:5,d} a day: rho {:.2f}, answer in {:6.1f} "
"min, P(wait > 1 h) {:.3f}".format(n, rho, w, p))
# escalation share 0.02: stable below 1,800 a day
# 600 a day: rho 0.33, answer in 15.0 min, P(wait > 1 h) 0.006
# 1,200 a day: rho 0.67, answer in 30.0 min, P(wait > 1 h) 0.090
# 1,500 a day: rho 0.83, answer in 60.0 min, P(wait > 1 h) 0.307
# 1,700 a day: rho 0.94, answer in 180.0 min, P(wait > 1 h) 0.677
# 1,750 a day: rho 0.97, answer in 360.0 min, P(wait > 1 h) 0.823
# 3,000 a day: rho 1.67, queue grows without limit
# escalation share 0.01: stable below 3,600 a day
# 600 a day: rho 0.17, answer in 12.0 min, P(wait > 1 h) 0.001
# 1,200 a day: rho 0.33, answer in 15.0 min, P(wait > 1 h) 0.006
# 1,500 a day: rho 0.42, answer in 17.1 min, P(wait > 1 h) 0.013
# 1,700 a day: rho 0.47, answer in 18.9 min, P(wait > 1 h) 0.020
# 1,750 a day: rho 0.49, answer in 19.5 min, P(wait > 1 h) 0.022
# 3,000 a day: rho 0.83, answer in 60.0 min, P(wait > 1 h) 0.307上交比例 2% 时,只有每天少于 1,800 场对话,队列才是稳定的。每天 600 场时,创始人三分之一的时间在忙,平均 15.0 分钟答复;每天 1,500 场时,平均答复时间 60.0 分钟,0.307 的上交要等一个多小时创始人才开始处理;1,750 场时是 360.0 分钟。答复时间按 上升,而不是与对话量成比例。把上交比例减半到 1%,稳定上限翻倍到每天 3,600 场,1,500 场时的答复时间降到 17.1 分钟。在一人公司里,稀缺的资源是创始人的注意力,杠杆是上交比例。
注
高利用率下的排队。 单服务台公式假定随机(泊松)到达与指数分布的处理时间。真实的上交会集中在上新与晚间高峰,在同样的平均负荷下,等待比公式所示更糟。
小结
误差沿智能体链条累积:每步正确率 0.95 的 10 步链条只交付 0.599。每步之后检查花费不多,却能抓住大部分错误。处理上交的创始人是一个排队系统,等待时间按 增长;决定一个人能承载多少对话的,首先是上交比例。
参考文献
Fan, Y. (2026). One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage: Coasean Transaction-Cost Collapse, Cyclic Script Retention and High-Margin Solopreneur Portfolios [一人公司自动化架构、数字人直播漏斗与任务级人工智能套利] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090235 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).