Under required delegation, subagents report a third of the conflicts they read
When three language models had to hand document review to subagents, both conflicting figures were read more often, but only 13.2 per cent of conflicts were detected in the pooled results (16.2 per cent with a single agent). Subagents reported both figures they had read in only 35.2 per cent of cases.
Researchers from UC San Diego, Intel and Hewlett Packard Enterprise published a preprint on arXiv on 17 November. It has not been peer reviewed. The study examines where conflicting information gets lost when a lead AI agent hands document review to subagents. A subagent reads its assigned documents in its own context window and returns a summary to the lead agent.
The researchers inserted 73 numerical conflicts into 11 due diligence tasks in Harvey LAB, a legal-work benchmark published by Harvey. For each conflict, they replaced a recurring figure with a different one in some of the folders. The agent had to write a risk report. The researchers’ tool tracked whether each figure appeared in the sources read, the subagents’ reports, the lead agent’s material and the final report.
The study tested GPT-5.6 Luna, GPT-6 Luna and Claude Haiku 5.5. Each model and workflow received the same dollar budget per task. There were three workflows: a single agent, optional delegation, and required delegation, in which the lead agent does not read sources itself.
The table shows pooled results for the three models. In the pooled results, even the best workflow surfaced fewer than a quarter of the conflicts. The language model claude-sonnet-5 judged the reports. The authors found three cases by hand in which it accepted an incorrect comparison.
| Workflow | Both figures read | Both figures with lead agent | Conflict detected |
|---|---|---|---|
| Single agent | 50.2 | 50.2 | 16.2 |
| Optional delegation | 62.8 | 54.8 | 24.2 |
| Required delegation | 68.7 | 24.4 | 13.2 |
Percentage of 438 observations per workflow. In the single-agent workflow, the lead-agent column shows the material the agent read itself. Source: Zhang et al., arXiv preprint.
Under required delegation, subagents read both figures in 301 cases but recorded them in their reports in only 106. In the authors’ interpretation, a subagent confined to its own portion does not recognise a figure as conflicting, because it cannot see the other document.
Detection depended on how the figures were split between subagents. When the same subagent read both figures, required delegation detected 34.2 per cent of conflicts. A single agent detected 22.5 per cent of the same cases. When the figures were split across different subagents, the shares were 10.0 and 18.9 per cent. The authors said the groups reflect what the agents themselves read and reported, and that the effect of the split cannot yet be separated from how hard the figures were to find.
The authors caution against splitting tasks that require cross-checking into rigid, non-overlapping parts. Under optional delegation, the lead agent may read sources and the subagents’ portions can overlap. This workflow detected the most conflicts with all models, although the GPT models used subagents in fewer than half of their runs. Because the experiment changed document partitioning and the lead agent’s reading restriction at once, the effect of required delegation cannot be attributed to either. The authors also note the limits of the data: there were only 11 tasks, the conflicts involved only figures, and each model was run twice.
Read more: How an agent works · Agent map