OpenAI single-agent LLM architecture reduces computational overhead relative to multi-agent orchestration in a...

Area of Science:

• Artificial Intelligence
• Robotics
• Space Exploration

Background:

• Mars rover missions necessitate advanced decision-support systems for autonomous operation under communication delays.
• Interpreting complex terrain, telemetry, and environmental data is crucial for mission success.

Purpose of the Study:

• To evaluate the effectiveness of multi-agent orchestration versus a single-agent baseline for simulated Mars rover decision support.
• To compare the performance and computational efficiency of different large language model (LLM) architectures (GPT-4o, GPT-5.5) in these scenarios.

Main Methods:

• A benchmark of 100 synthetic Mars rover scenarios was used to compare single-agent and multi-agent architectures.
• Performance metrics included decision accuracy, hazard identification (F1 scores), error counts, latency, and token usage.
• Statistical analyses, including Holm-Bonferroni adjustment, were applied to scenario-level comparisons.

Main Results:

• Single-agent architectures showed a numerical advantage in decision accuracy and hazard alignment, though not always statistically significant.
• Multi-agent orchestration generated more comprehensive hazard lists but did not reliably improve aggregate decision accuracy or F1 scores.
• The single-agent architecture demonstrated significantly lower latency and token usage, indicating greater computational efficiency.

Conclusions:

• For short-context, tool-less decision-support tasks, multi-agent orchestration is a cost-bearing design choice, not an inherent improvement.
• The study provides a benchmark for assessing the operational cost-effectiveness of LLM-based orchestration in mission-inspired workflows.