Why Enterprise Search Agents Need Multi-Turn Reinforcement Learning

The Cost of Reliability in AI Agents

Enterprises are increasingly deploying AI agents to handle complex information retrieval tasks. Unlike traditional search tools that require users to craft precise queries, these agents autonomously decide what to search for, which strategies to employ, and when to stop. They operate across multiple rounds of interaction, refining their approach based on previous results.

However, a significant barrier remains: getting these multi-step behaviors to work reliably is difficult. Base language models do not inherently know an organization’s specific tools or environment. Prompting small models rarely yields dependable multi-turn behavior, while prompting frontier (largest) models often works but incurs high latency and cost. For many businesses, this trade-off between speed and capability is unsustainable at scale.

Refer & earn: 30% of a client’s first month, then 10% of every month after that, for up to 24 months per client. See the partner programme →

The Limitations of Traditional Fine-Tuning

Fine-tuning offers a third path, teaching a small model your specific tools and environment directly. This promises the speed and cost benefits of a small model with the reliability typically associated with frontier models. Yet, traditional fine-tuning methods fall short for agentic workflows.

Supervised fine-tuning (SFT) depends on expert demonstrations of ideal multi-turn trajectories. These demonstrations are costly to collect and often do not exist for unique enterprise setups. Single-turn reinforcement learning (RL), such as RL with verifiable rewards (RLVR), scores one response at a time. But a search agent makes interdependent decisions across many turns. Optimizing a single step in isolation misses the dependencies between steps, leading to suboptimal performance.

Multi-Turn Reinforcement Learning (MTRL)

The solution lies in optimizing the agent across the full multi-turn trajectory. Multi-turn reinforcement learning (MTRL) addresses this by framing an agentic task as a sequence of decisions. It uses multi-turn rollouts to generate training data and optimizes the model with policy gradient algorithms. The reward signal only needs to reflect whether the final outcome was good, allowing the agent to learn effective decision-making across the entire sequence.

This approach is particularly suitable for search agents because it provides a clear reward signal (retrieval quality), a multi-turn interaction loop, and a well-defined environment consisting of search tools. By running a smaller, specialized model, organizations achieve faster, cheaper inference without sacrificing the reliability required for critical business operations.

Real-World Results from AWS

A recent case study published by Amazon Web Services demonstrates the practical application of this technology. Using Amazon SageMaker AI MTRL, researchers fine-tuned a Qwen3.6-27B model for an enterprise search setting. The agent was equipped with two specific tools and limited to a maximum number of turns to prevent excessive responses and encourage efficiency.

The training setup focused on three key components: datasets, reward functions, and configuration. The primary metric used was nDCG@10 (Normalized Discounted Cumulative Gain at rank 10), a standard information retrieval metric measuring how well retrieved documents match ideal rankings. This metric served directly as the reward function. Additionally, a penalty of -1 was assigned if the agent reached maximum turns or token limits, explicitly teaching the model to avoid these failure modes.

The results were substantial. MTRL fine-tuning improved the search agent on three of four held-out benchmarks. The largest gains appeared on BrowseComp-Plus, with a 23.7 percent increase in nDCG@10, and WixQA, with an 18.4 percent gain. More strikingly, the reliability improvements were dramatic. On BrowseComp-Plus, the failure rate fell from 22.89 percent to just 0.68 percent.

This reduction in failure rates indicates that the agent learned not only to search better but also to complete tasks within its turn and token budget. The training curves showed steady increases in reward on both training and validation sets, saturating by the end of training, which suggests that further training would not yield additional benefits.

Implications for Business Decision-Makers

For IT leaders and operational managers, these findings highlight a shift in how AI infrastructure should be evaluated. The assumption that larger models are always necessary for complex reasoning tasks is being challenged by specialized fine-tuning techniques.

Cost Efficiency: By fine-tuning smaller models like Qwen3.6-27B, organizations can significantly reduce inference costs compared to using frontier models. This is crucial for applications requiring frequent, real-time interactions.

Reliability and Control: The drop in failure rates demonstrates that MTRL can produce agents that adhere to operational constraints. In enterprise settings, where hallucinations or infinite loops can disrupt workflows, this predictability is valuable.

Implementation Simplicity: One of the appeals of Amazon SageMaker AI MTRL is the minimal configuration required. Hyperparameters can be left at default values, reducing the need for deep reinforcement learning expertise. This lowers the barrier to entry for teams looking to customize AI models for specific business environments.

What to Check Before Acting

If your organization is considering implementing AI agents for search or other multi-step tasks, several factors warrant attention before committing resources.

First, assess your current tooling and environment. MTRL requires a well-defined environment with clear reward signals. If your search tools or data sources are inconsistent, defining a reliable reward function may be challenging. Ensure you have a robust way to measure retrieval quality, such as nDCG@10, to guide the training process.

Second, evaluate your existing model capabilities. While fine-tuning improves performance, starting with a capable base model is essential. The AWS example used Qwen3.6-27B, a model supported in specific regions like US West (Oregon). Verify that your chosen provider supports the necessary infrastructure and that your team has access to the required compute resources.

Third, consider the operational overhead. Although MTRL simplifies some aspects of RL, training can span multiple days. Implementing checkpoint resume support and managing resource cleanup are necessary to avoid ongoing charges and ensure efficient development cycles. Teams must be prepared to monitor training curves and adjust configurations as needed.

Finally, align AI initiatives with business outcomes. The goal is not just technical improvement but enhanced productivity and user satisfaction. Measure the impact of your agents on key metrics such as response time, accuracy, and user engagement. Use these insights to refine your approach and justify further investment in AI customization.

Conclusion

Multi-turn reinforcement learning represents a significant advancement in making AI agents practical for enterprise use. By fine-tuning smaller models with MTRL, organizations can achieve the reliability of frontier models at a fraction of the cost and latency. As demonstrated by AWS, this approach yields substantial improvements in retrieval quality and reduces failure rates dramatically.

Businesses that adopt this strategy will likely find themselves better positioned to scale AI-driven processes efficiently. The key is to start with clear objectives, robust evaluation metrics, and a willingness to experiment with specialized fine-tuning techniques. As the technology matures, we can expect more accessible tools and frameworks to emerge, further democratizing the power of agentic AI.

Sources

Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI | Amazon Web Services