# Dorsal UX vs Codex web search

Chris Benson's adversarial agent swarm experiment - November 18, 2026.

**Dorsal’s reply answered the specific question. Mine missed Benson’s experiment and substituted broader examples.**

| Aspect | My web-only reply | Dorsal UX reply |
|---|---|---|
| **Benson’s experiment** | Couldn’t find it. Described his general discussions of swarms and cyber defense. | Found him experimenting with opposing agent roles to improve enterprise architecture. |
| **Closest engineering example** | Multiagent debate for reasoning and factual accuracy. | AI Jason’s five-agent debugging: competing hypotheses, shared findings, debate, and final reports. |
| **Other use cases** | Scientific discovery, vulnerability exploitation, parallel research. | Verified parallel coding, strategic review, vulnerability remediation, product evaluations, independent forecasting. |
| **Practical detail** | Explained broad patterns and published results. | Identified concrete roles and workflows: workers, challengers, verifiers, and independent predictions. |
| **Sources** | Direct links to relevant primary sources, but none established the requested Benson experiment. | Eight source links, including the decisive Benson transcript record and closely related demonstrations. |
| **Qualifications** | Acknowledged that I hadn’t found the experiment. | Distinguished experiments from production results, thematic similarities from recorded connections, and benefits from coordination costs. |
| **Recorded performance** | Sol 6.1 at xhigh effort. 3 min 35 s from request to final reply; given a 120-second allowance. | **37 seconds, 6 Dorsal calls, 8 full records fetched.** |

**The substantive advantage was discovery:** Dorsal surfaced the exact experiment and more applicable engineering examples. My reply provided credible background but failed the central lookup.

## Scope

This is an informal comparison of two actual responses, not a controlled benchmark. Both used Sol 6.1; the web-only run used xhigh effort in Codex. The prompts and retrieval capabilities differ. The web-only session record establishes 214.822 seconds from the request at 21:42:15.712 UTC to the final reply at 21:45:50.534 UTC. Dorsal's 36.985-second research duration is corroborated by the recovered browser session and Vercel metrics. These durations cover different boundaries and are not a controlled speed comparison. Dorsal's effort setting and either run's token consumption were not recovered. The example establishes what each run found; it does not isolate graph traversal or establish better outcomes from using agent swarms.

Full replies, prompts, sources, and recovered trace: [full replies and trace](https://godorsal.com/#benson-agent-swarms).
