AI Agents' Correlated Errors Threaten Consensus: New EBFT Model
Even honest AI validators can form quorums that certify invalid transitions due to shared reasoning flaws.
Jun He and Deying Yu's new paper, 'The Honest Quorum Problem: Epistemic Byzantine Fault Tolerance for Agentic Infrastructure' (arXiv:2607.16109), tackles a critical blind spot in distributed systems powered by AI agents. Traditional Byzantine Fault Tolerance (BFT) assumes that any participant not actively malicious will correctly execute protocol semantics. But agentic validators—AI models acting as nodes—can be protocol-compliant yet still endorse invalid state transitions due to reasoning errors, which the authors call epistemic faults. Worse, because agents often share model weights, training data, prompts, or toolchains, these mistakes are highly correlated. An 'honest quorum'—a majority of honest (protocol-following but semantically wrong) agents can form a certificate for an invalid transition, breaking safety without any Byzantine behavior. This is the Honest Quorum Problem: agreement alone no longer guarantees semantic validity.
The authors propose a new fault-tolerance model, Epistemic Byzantine Fault Tolerance (EBFT), that augments the classic Byzantine fault bound with two separate, confidence-indexed quantities. eδ bounds the number of coherent invalid endorsements coming from outside the traditional Byzantine set, while uε bounds unusable validator support that degrades liveness. These quantities decouple semantic safety risk from liveness degradation. The paper derives quorum-threshold conditions for semantic validity, consensus agreement, liveness, and feasible threshold selection, along with a calibration methodology for estimating these budgets. A key finding: simply adding more nominally distinct agents does not improve fault tolerance unless it measurably reduces the upper-tail concentration of invalid endorsements or unusable support. For professionals deploying agentic infrastructure, this work provides a rigorous framework to quantify and mitigate risks from correlated AI reasoning failures, moving beyond the binary 'honest vs. Byzantine' mindset.
- Identifies a new failure mode: epistemic faults where protocol-compliant AI agents endorse invalid transitions due to reasoning errors
- Introduces EBFT with two independent budgets: eδ for coherent invalid endorsements and uε for unusable support, decoupling safety from liveness
- Shows that adding more agents only increases fault tolerance if it reduces upper-tail error concentration, not merely nominal diversity
Why It Matters
As AI agents become validators in distributed systems, we can no longer assume compliance equals safety—correlated errors demand new fault-tolerance models.