This thesis examines how different degrees of agent autonomy influence task efficiency, answer quality, and emergent behaviours in LLM-based systems built for marketing automation. Two architectures are compared within the same production platform. The first is an agent-orchestrator design in which one general-purpose agent reasons over a unified tool set and support subagents. The second is a multi-agent design in which platform-specific specialist agents activate from connected data sources, each independently decides whether the query warrants its attention, and all communicate findings through a shared message board before a synthesis step produces the final response. Evaluation uses 50 marketing scenarios, an n = 5 repeated report design, and combines trace-grounded agent-as-a-judge scoring, platform coverage, wall-clock latency, token usage, tool use, and estimated model cost. Across completed runs, the agent-orchestrator remained faster and cheaper, averaging 48.0 seconds and $0.3171 per run. The multi-agent system averaged 85.6 seconds and $0.8617, but produced higher-quality answers across 208 matched judged comparisons: it won 149, while the agent-orchestrator won 44 and 15 were ties. Mean composite judge score improved by +0.51 points on the 1-5 scale, with the strongest gains in completeness and actionability. The results indicate that agent-orchestrator execution is preferable for narrow, precise retrieval tasks, whereas multi-agent execution is more robust when the task requires exploration, synthesis, and action planning across several marketing data sources.