Accelerate Agentic Tool Calling with Serverless Model Customization in Amazon SageMaker AI

Smarter AI Agents with Serverless Customization

Agentic AI systems are transforming how we interact with technology, but they often stumble when trying to call tools, leading to issues like "hallucinating" tools, passing incorrect parameters, or failing to ask for clarification. These common pitfalls can erode user trust and hinder production deployment. Amazon SageMaker AI is tackling these challenges head-on with its new serverless model customization capabilities, designed to significantly improve the accuracy and reliability of agentic tool calling.

This innovative feature allows developers to fine-tune large language models (LLMs) without the typical operational overhead. A key technique at the heart of this advancement is Reinforcement Learning with Verifiable Rewards (RLVR). Using RLVR, models learn to generate more accurate and appropriate tool calls by receiving feedback on their candidate responses. SageMaker AI supports a wide range of model families, including Amazon Nova, GPT-OSS, Llama, Qwen, and DeepSeek, and offers various customization techniques beyond RLVR, such as Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning from AI Feedback (RLAIF).

A practical demonstration involved fine-tuning Qwen 2.5 7B Instruct for tool calling using RLVR. The results were impressive: the fine-tuned model improved its tool call reward by a remarkable 57% over the base model when evaluated on unseen scenarios. This significant leap forward means AI agents can now be more dependable and efficient in tasks requiring external tools, from querying databases to triggering complex workflows. You can explore the technical details and implementation steps for this feature on the AWS Machine Learning Blog.

Why RLVR Matters for Agentic Workflows

Traditional reinforcement learning (RL) approaches for model customization often come with substantial operational burdens. Managing GPU procurement, memory orchestration between rollout and training phases, reward infrastructure, checkpointing, and hyperparameter sensitivity can quickly become complex and time-consuming. SageMaker AI's serverless model customization removes this infrastructure management, allowing AI developers to focus entirely on their models, data, and reward functions.

RLVR works by having the model generate multiple candidate responses for each prompt – typically eight in the example use case. These candidates are then evaluated by a meticulously designed reward function that verifies their correctness. The model updates its policy to favor the responses that worked best using Group Relative Policy Optimization (GRPO). GRPO compares each candidate's reward score against the mean score of the group, reinforcing responses that score above average. Over time, the model effectively learns the precise format for tool calls and, crucially, when to invoke a tool versus when to ask for clarification.

This distinction is vital for production-ready agents. For instance, an agent needs to learn the difference between a clear request like "Get weather for San Francisco" (which requires a tool call) and a vague one like "Get the weather" (which requires a clarifying question). This refined decision-making process is a core benefit of RLVR in SageMaker AI, leading to more robust and context-aware agents. Dive deeper into the benefits and mechanisms of this serverless approach by visiting the AWS Machine Learning Blog.

Getting Started with Serverless Customization

To begin leveraging serverless model customization in Amazon SageMaker AI, there are a few prerequisites. You'll need an active AWS account, an AWS IAM role with the necessary permissions, a SageMaker AI domain with Studio access, and an Amazon S3 bucket for data storage. Once these are in place, the process for fine-tuning a model is streamlined.

Preparing your training data is a crucial step. The demonstration utilized 1,500 synthetic examples generated using Kiro across five practical tool schemas: weather, flights, translation, currency conversion, and statistics. These examples were carefully distributed to teach the agent three distinct behaviors: Execute (60% of examples, where the user provides all required parameters for a tool call), Ask for clarification (where required parameters are missing), and Refuse (for harmful or out-of-scope requests). This comprehensive dataset ensures the model learns to handle a variety of real-world scenarios.

Training and validation metrics are seamlessly tracked through integrated MLflow, providing clear visibility into the model's learning progress. By selecting your desired foundation model, configuring RLVR as the customization technique, and pointing to your prepared data and reward function, SageMaker AI handles the underlying infrastructure, allowing you to focus on developing high-performing, reliable agentic AI applications.

Read more: Accelerate Agentic Tool Calling with Serverless Model Customization in Amazon SageMaker AI to dive deeper into the technical implementation and start building more reliable AI agents today.