Agents user guide

Selecting the right model defines how your agent behaves in real scenarios. The model influences accuracy, stability, speed, and cost.

Choosing a model for an agent is an iterative process, not a one-time decision. You choose an initial model during design so you can build and test the agent, and later refine that choice once evaluations show how different models behave with your prompts, tools, data, and failure scenarios. Evaluations often show that a lower-cost model meets the same quality requirements as a more expensive option, or that a different model performs more reliably on specific edge cases.

This section helps you understand model settings, apply best practices, and run the steps needed to choose the most effective and cost-efficient model for your use case.

Model settings control how the underlying AI model produces outputs. Two settings have the most impact on agent performance: model and temperature.

The model choice affects capability, latency, cost, and specialized strengths. Different models excel at tasks such as reasoning, coding, or summarization. To understand which one fits your agent’s workload, run evaluations that compare how models perform with your actual prompts and scenarios, as described in the following sections.

Temperature controls the randomness and creativity of the model's responses. Typical ranges include:

During design, select a model that broadly fits your agent’s workload. This initial model serves as the baseline you use to build prompts, integrate tools, and test flows. As a best practice, we recommend you start with a general-purpose or lower-cost model and expect to revisit this choice after evaluations.

For the full list of models supported in UiPath Automation Cloud and their regional availability, see Model availability and routing.

Temperature controls how deterministic or variable a model’s responses are. In most enterprise agents, consistency is more important than creativity.

Temperature should be tuned sparingly and always validated through evaluations.

Evaluations are where model selection becomes evidence-based.

First, start with a working agent. Run it in debug mode with different inputs across your key scenarios to confirm the full flow behaves as expected. Once the agent works end to end, build your evaluation set from real runs, either by using Add to evaluation set directly from a debug run or by downloading runtime runs and importing them into your evaluation set.

Build evaluation sets using:

Avoid relying only on synthetic or auto-generated cases, which can overestimate real-world performance.

Run the same evaluation set across multiple models and configurations. At this stage, you can decide whether differences in quality justify differences in cost, latency, or stability. Running the same scenarios across configurations makes these trade-offs visible. It is common and expected to change the selected model after reviewing evaluation results.

To configure and compare different model settings within an evaluation set:

To make comparisons fair:

Each model added to your evaluation set triggers a new run, and you can review the results for each run in the Results table. When reviewing results, you are not just looking for the highest score, you are deciding which trade-offs matter most for your agent.

Review the evaluation results to understand how each configuration performs. Look for:

Do not select a model based only on an average score. When reviewing evaluation results, consider:

A slightly lower-scoring model may be preferable if it is significantly cheaper and more stable.

The steps below summarize the core process described in this section. Use them as a quick reference when selecting and refining a model for your agent: