TLDR: Routine is a novel framework that provides structured plans for AI agents, significantly improving their ability to perform multi-step tasks and use tools reliably in enterprise settings. It enhances execution accuracy for models like GPT-4o and Qwen3-14B and allows smaller models to achieve high performance through specialized training and knowledge distillation, making AI agent deployment more practical and stable in real-world business scenarios.
Deploying AI agent systems in a business environment often comes with significant hurdles. Common AI models frequently lack the specific knowledge needed for domain-specific processes, leading to disorganized plans, overlooking crucial tools, and unstable performance. This can make it difficult for companies to truly leverage the power of AI for automating complex tasks.
To tackle these challenges, a new framework called Routine has been introduced. Routine is designed as a multi-step agent planning framework that brings much-needed structure, clear instructions, and smooth parameter passing to guide an AI agent’s execution. This allows agents to perform multi-step tasks involving various tools with high stability and accuracy.
The impact of Routine has been quite impressive in real-world enterprise scenarios. For instance, the execution accuracy of GPT-4o, a powerful AI model, saw a dramatic increase from 41.1% to 96.3% when guided by Routine. Similarly, Qwen3-14B, another large language model, improved its performance from 32.6% to 83.3%. These results highlight Routine’s effectiveness in making AI agents more reliable for business operations.
Beyond just guiding execution, Routine also plays a crucial role in training AI models. Researchers created a training dataset that follows the Routine framework and used it to fine-tune Qwen3-14B. This resulted in an accuracy increase to 88.2% in scenario-specific evaluations, showing that models can better adhere to execution plans when trained with Routine’s structured guidance.
Furthermore, Routine-based data distillation was used to create a specialized dataset for multi-step tool-calling in specific business scenarios. Fine-tuning models on this distilled dataset boosted their accuracy significantly. For example, Qwen3-14B’s accuracy reached 95.5%, almost matching the performance of GPT-4o. This demonstrates Routine’s capability to help AI models learn domain-specific tool-usage patterns and adapt to new situations, making them highly effective for enterprise deployment.
The Routine framework is built around four core modules of an AI agent system: Planning, Execution, Tools, and Memory. The Planning Module, often assisted by domain experts, generates the structured Routine. The Execution Module, typically a smaller, specialized AI model, follows this Routine to make tool calls. The Tool Module, using a system like MCP servers, provides the actual tools and their definitions. Finally, the Memory Module manages both long-term ‘Procedure Memory’ (storing Routines) and short-term ‘Variable Memory’ (handling intermediate results and parameters) to ensure efficient and accurate task completion.
This approach significantly reduces the computational load and improves accuracy by providing clear, step-by-step instructions, rather than relying solely on the AI model’s autonomous reasoning for complex tasks. It also allows for the use of smaller, more efficient models for execution, making AI agent systems more practical and cost-effective for real-world enterprise deployment.
Also Read:
- AgentFly: A New Framework for Training Language Model Agents with Reinforcement Learning
- LLMs Learn to Think Smarter with Hierarchical Budget Policy Optimization
In essence, Routine offers a practical and accessible method for building stable AI agent workflows, accelerating the adoption of these systems in businesses, and advancing the vision of AI for process automation. For more in-depth information, you can refer to the original research paper: Routine: A Structural Planning Framework for LLM Agent System in Enterprise.
