The proliferation of agent orchestration frameworks like LangGraph, CrewAI, and Google ADK has centered on an external orchestrator managing LLM state and routing. However, a controlled comparison detailed on arXiv reveals a simpler, more effective paradigm for procedural tasks: embedding the entire procedure directly into the system prompt, allowing the LLM to self-orchestrate.
Self-Orchestration Dominates Procedural Tasks
Across travel booking, Zoom technical support, and insurance claims processing, the in-context approach consistently outperformed external orchestration. For a 55-node insurance claims task, the in-context method achieved scores of 4.53-5.00 on a 5-point scale. In contrast, a LangGraph orchestrator using the same LLM yielded scores between 4.17-4.84. This agent orchestration frameworks comparison highlights a critical performance gap.
Reduced Failures with In-Context Control
The practical implications are stark. The orchestrated system failed in 24% of travel booking conversations, compared to just 11.5% for the in-context baseline. For insurance claims, failures dropped from 17% to 5%. This demonstrates that advances in frontier model capabilities have rendered external orchestration superfluous for multi-turn conversations with defined procedures, marking a significant shift in agent design. The researchers observed a notable shift in reliability, with the in-context approach demonstrating superior robustness.
