The Agent Telephone Problem: How Small Errors Compound When AI Agents Delegate to Each Other
Year: 2026
Type: preprint
Abstract: An agent that is right 95 percent of the time sounds reliable. Chain fourteen of them together, each handing its result to the next, and the final answer is right about 49 percent of the time: a coin flip. Nobody in the chain is careless, nobody is broken, and yet the system as a whole has quietly become a guessing machine. This article works out, for one deliberately simple model of such a chain, three things: how fast reliability decays, where a limited budget of verifying checks buys the most reliability back, and when copying the chain and voting helps or hurts. Two scope notes belong up front. First, this is a stylized simulation study: every number below is computed from the model, by exact formula and by Monte Carlo simulation, and none of it is a measurement of any deployed LLM agent system. Second, the compounding itself is not claimed as a discovery; it is multiplication, and it has been measured in real agent systems by others (see Related work). The contribution here is narrower: a placement and cost analysis of correcting verifiers, stated exactly enough to be checked. One finding cuts against the obvious remedy: in the tested setting where checks also damage clean work at a high rate (a 10 percent false-positive rate), checking every hop performs worse than a single end check, so within this model more checking is not automatically better.
Author: Pande, Rishabh
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7229797311
