Zenaique

Why a 90 percent per step correct multi-agent chain is not 90 percent correct overall

Flashcard·Medium·4.0 · 0·~30s·Asked atLocusSwiggyUnity
Attempt it
TL;DR

Errors compound multiplicatively, so 0.9 to the sixth is about 0.53; chain length is exponential cost, and cutting hops or adding a final critic beats grinding up per-step accuracy.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Imagine a relay where six people in a row each have to copy a sentence by hand, and each person gets it right nine times out of ten. It feels like the message should arrive almost perfectly, because every person is so good. But the only way the final sentence is right is if all six people happened to get it right that round. Multiply 0.9 by itself six times and you get about 0.53. The message arrives correctly only a little better than half the time. That is why teams that look strong at every station can still fail more than half their tasks. Shortening the relay or adding one person who checks the final sentence against the original both help a lot more than trying to make each person slightly more accurate.

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Multi agent chains have a deceptively simple failure mode: errors compound. The whole task succeeds only when every hop succeeds, and the probabilities multiply. That fact alone explains a large share of the disappointment that production teams experience when they ship a multi agent architecture that looked great in demos.

This question gives you the smallest possible version of the calculation: six agents, each 90 percent correct, and asks for the end to end rate. The number is 53 percent. The lesson is what to do about it.

The multiplicative formula and the shape of the curve

For a chain where each step succeeds independently with probability p, and the task succeeds only when all steps succeed, end to end accuracy is the product of per step probabilities. For a uniform chain of length n at rate p, this collapses to:

Pend to end=pnP_{\text{end to end}} = p^n

Plug in p equal to 0.9 and n equal to 6 and you get about 0.531. That is the number that surprises engineers who reason additively. 'Six steps that are each 90 percent good must be roughly 90 percent good overall' is wrong by almost 40 percentage points.

The slope is the headline

The interesting thing is not the value at n equal to 6; it is how fast the value moves as n changes.

  • n equal to 4: 0.656
  • n equal to 6: 0.531
  • n equal to 8: 0.430
  • n equal to 10: 0.349

Each additional hop costs roughly 5 to 6 percentage points of end to end accuracy at the 90 percent per step rate. That is the exponential biting. A team that decides to 'add a planning agent and a routing agent and a formatting agent' to a previously working chain is paying a real reliability tax.

Why measured reality is usually worse than the floor
Three levers, ranked by leverage
When the multiplicative frame does not apply
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.
python
# A back of envelope tool: end to end accuracy under multiplicative compounding
from functools import reduce
from operator import mul

def chain_accuracy(per_step):
    return reduce(mul, per_step, 1.0)

uniform_chain = [0.9] * 6
print(round(chain_accuracy(uniform_chain), 3))  # 0.531

# Lever 1: cut hops
shorter = [0.9] * 4
print(round(chain_accuracy(shorter), 3))  # 0.656

# Lever 2: push per step accuracy up
stronger = [0.95] * 6
print(round(chain_accuracy(stronger), 3))  # 0.735

# Lever 3: add a terminal verifier that recovers some fraction of failures
base = chain_accuracy([0.9] * 6)
verifier_recovery = 0.6  # critic catches 60% of compounded errors
recovered = base + (1 - base) * verifier_recovery
print(round(recovered, 3))  # 0.812
InterventionEffect on a 6 hop 90 percent chainCost
Baseline (6 hops, 0.9 each)0.531 end to endSix agent calls
Cut to 4 hops0.656Lose two agents, redistribute work
Per step 0.9 to 0.950.735Heavy per agent prompt and eval work
Add terminal critic (60 percent recovery)0.812One extra agent call per task
Cut to 4 hops + add critic0.862Structural redesign plus one call

Real products, models, and research that use this idea.

  • Anthropic's 2024 multi agent research blog described 'compounding error' as the dominant failure mode in long agent chains, motivating shorter trajectories with verification hops.
  • LangChain's 2025 multi agent design guidance recommends fewer than four sequential hops and a terminal critic for production reliability above 80 percent.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would the analysis change if you modelled correlated errors instead of independent ones?
A

Introduce a correlation coefficient between adjacent agents; the joint failure probability becomes higher than the product, so end to end accuracy drops below the multiplicative floor. Use a small Monte Carlo to estimate the gap on a real workload.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Reasoning additively about chain accuracy, so a team of six 90 percent agents feels like it should be 'about 90 percent good' instead of 53 percent good, and then debugging individual agents when the architecture itself is the problem.

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • The multiplicative formula for end to end success in a chain

  • Why six 90 percent agents land near 53 percent, not 90 percent

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
Why AutoGen 0.4 makes TerminationCondition a first class primitive instead of leaving it to convention
Flashcard·Medium