AI agents sharpen tool calling with SFT and DPO on SageMaker

Tool-calling accuracy just got a major boost.

Amazon SageMaker AI training jobs are being used to train small language models with a two-pronged approach: supervised fine-tuning, which builds a high-quality dataset aligned with intended tool use, and direct preference optimization, which nudges outputs toward preferred tool interactions. The team reports that this combination helps the model learn tool-specific language, commands, and constraints more reliably, improving how it chooses and formats when to call a tool.

By comparing a base model to several fine-tuned variants, the evaluation centers on tool-calling accuracy and the ability to maintain correct workflow sequences. The paper shows that pairing SFT with DPO lets the model better align its actions with the actual tools available, reducing miscalls and malformed parameters that often derail multi-step tasks. Benchmarks indicate the tuned variants outperform the baseline on key measures of tool usage, suggesting a more dependable handoff between language reasoning and tool execution.

From a practitioner standpoint, the approach clarifies several engineering realities. Data quality is paramount: SFT relies on a curated dataset that mirrors real tool usage, so the model can internalize tool-specific language and constraints. Poor or biased data can lead to brittle tool calls that break when tools change or expand. The team reports that DPO introduces a further optimization layer, optimizing for preferred behaviors like correct parameter syntax and appropriate sequencing, but it also requires careful construction of preference data to avoid overfitting to narrow patterns.

There are clear tradeoffs to manage in production. Fine-tuning adds compute and data costs, and teams must maintain updated datasets as tools evolve, which can be resource-intensive. Yet the payoff is tangible: more reliable automation translates to faster task completion, fewer user-visible errors, and lower support costs as agents handle more steps autonomously. The work also underscores the importance of robust evaluation; measuring tool-calling accuracy across diverse tool sets and failure modes helps prevent overconfidence in a narrow set of scenarios.

The study points to concrete next steps to watch for. Scaling the approach to larger tool sets and cross-domain tasks will test generalization, while evolving environments will demand ongoing retraining to keep tool calls aligned with current capabilities. In short, the SFT plus DPO recipe offers a practical path from pilot experiments to production-grade agents, provided teams invest in high-quality data curation, thoughtful preference design, and disciplined monitoring of tool-call outcomes.

In a field where automation often promises efficiency but struggles with reliability, the AWS-focused method provides a clear engineering blueprint: curate strong tool-specific data, couple supervised signals with human-preference guidance, and rigorously compare baselines to tuned variants. The outcome is not just better tool selection, but a smoother, more predictable workflow that keeps automation aligned with real-world tool capabilities, and provides a framework for teams to iterate without reinventing the wheel.