Benchmark Finds Delegation Hides Conflicts in AI Agents
A benchmark titled "Multi-Agent Memory Inconsistency: Where Factual Conflicts Disappear in LLM Agents" was indexed on November 17, 2026, with an update posted the following day, documenting a measurable gap between what multi-agent systems retrieve and what they actually flag as contradictory (alphaXiv).
What the Benchmark Measured
The study built controlled factual conflicts into Harvey LAB document-review tasks, a setting designed to mimic the kind of legal and compliance document work increasingly handed to AI agents (pith.science). Researchers compared two setups: a single agent handling retrieval directly, versus a delegated structure where worker agents split the task and reported back to a coordinating agent.
The headline numbers, pooled across three language models, show the trade-off plainly. When retrieval was delegated to worker agents rather than handled by a single agent, the rate at which both conflicting versions of a claim were retrieved rose from 50.2% to 68.7% (mindpattern.ai). But the rate at which the system actually detected and surfaced the conflict fell from 16.2% to 13.2% over the same comparison (theagentwatch.com).
Why the Gap Appears
The pattern suggests that splitting work across agents makes it more likely that both sides of a factual disagreement end up somewhere in the system's combined memory — yet less likely that any single agent, or the coordinator synthesizing their outputs, notices the contradiction exists. More raw information is surfaced, but less of it is reconciled. The indexed record does not disclose the authors' names, institutional affiliations, or the specific three models tested, and the available sources are repository-indexed summaries rather than a peer-reviewed publication from a named lab or a wire-service report.
Why It Matters Now
The findings land as document-review and legal-research workflows are among the most commonly cited enterprise use cases for delegated agent systems, where a coordinator agent assigns subtasks to specialized workers. If conflict detection degrades precisely when delegation increases retrieval coverage, systems built on multi-agent architectures may be quietly less reliable at catching contradictory facts than single-agent setups, even though they appear to pull in more relevant material. The benchmark offers no fix, only a measured gap between retrieval and detection that future agent-memory designs will need to close.


