The Agent Telephone Problem: How Small Errors Compound When AI Agents Delegate to Each Other
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. This article measures that collapse with a small, fully reproducible simulation (20,000 seeded runs per configuration, standard-library Python). At a 5 percent error rate per handoff, end-to-end reliability falls from 0.78 after 5 agents to 0.36 after 20 and 0.21 after 30. A verifier placed at the end of a 20-agent chain restores reliability to 0.97 at the cost of a single check per run, while the same verifier placed in the middle reaches only 0.58: placement beats quantity. The cheapest placement that restores reliability above 0.90 depends on checker quality: one end check for a strong checker (catch rate 0.95), checks every 5 hops for a weaker one (0.80). Majority-vote redundancy over parallel chains amplifies rather than repairs: chains that are individually right 36 percent of the time vote their way down to 0.29 (3 chains) and 0.25 (5 chains), and a 2 percent common failure mode caps what redundancy can achieve at any scale. The article closes with four design rules for multi-agent systems and a plain statement of the model's limitations.
Authors
- Rishabh Pande
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24235779
- Primary Topic
- Distributed systems and fault tolerance
- Type
- preprint