In-Context Prompting Obsoletes Agent Orchestration for Procedural Tasks This paper investigates whether the complex frameworks currently used to manage AI agents—such as LangGraph or CrewAI—are actually necessary for completing procedural tasks. These frameworks typically use an "orchestrator" to track conversation state and dictate the agent's next move at every step. The authors conduct a controlled study to test if a simpler approach, where the entire procedure is provided directly in the system prompt, allows a frontier model to manage itself more effectively than an external orchestrator. The Shift from Orchestration to Self-Management Current agent development often relies on "orchestration," where an external system breaks a workflow into individual nodes and injects specific instructions at each turn. The authors argue that this architecture is often counterproductive. By contrast, their "in-context" approach provides the model with the full, serialized flowchart of the procedure within the system prompt. This allows the model to maintain a holistic view of the conversation, enabling it to handle transitions, decision-making, and state tracking internally without needing an external controller to intervene at every step. Comparing Performance Across Domains The researchers tested both methods across three distinct domains: travel booking, Zoom technical support, and insurance claims processing. Using 1,200 total conversations, they evaluated performance based on task success, information accuracy, consistency, graceful handling, and naturalness. The results were consistent across all tests: the in-context approach outperformed the orchestrated system in every metric. The orchestrated systems frequently suffered from "fragmented reasoning," where the model struggled to maintain coherence because its instructions were broken up into isolated, node-by-node segments. Reliability and Failure Modes A significant finding of the study is that orchestration introduces failure modes that do not exist in the in-context baseline. The orchestrated systems failed on 9% to 24% of conversations, depending on the domain, compared to only 0.5% to 11.5% for the in-context approach. These failures often manifested as the agent looping indefinitely, skipping necessary steps, or prematurely ending the conversation. Because the orchestrated agent relies on external routing calls at every decision hub, any error in the routing logic propagates downstream, causing the entire process to collapse. Efficiency and Practical Considerations While the in-context approach is more reliable and produces higher-quality interactions, it does come with a different resource profile. Because the entire procedure is included in every API call, the in-context method consumes more tokens than the orchestrated version. However, the orchestrated approach requires more total API calls to manage routing, which adds latency and increases the overall cost per conversation. Ultimately, the authors conclude that for procedural tasks, the "bottleneck" is not a lack of better orchestration, but the orchestration architecture itself. As frontier models continue to improve, the need for complex external scaffolding is diminishing.