Autopilot agent

Documentation for the Autopilot agent experience, a multi-step conversational interface available in Studio Desktop and Test Manager.

Autopilot agent is the multi-step conversational experience that lets you interact with Autopilot through a chat interface inside UiPath products. Unlike Autopilot actions, which are single-trigger operations, the agent maintains context across a conversation and can perform work one step at a time with you driving each step.

Autopilot agent is available in Studio Desktop (2025.10.8+), Studio Web, and Test Manager. In Studio Desktop, it has been succeeded by the full coding agent experience starting with STS 2026.0.195+, but the agent remains supported. Refer to Product availability for the full list of products and capabilities.

How it works

The Autopilot chat interface provides a conversational panel where you can:

• Ask questions about your project, automations, or UiPath documentation
• Request workflow generation, editing, or explanations
• Iteratively refine results through follow-up prompts
• Get contextual suggestions based on the product and page you are working in

Each product where Autopilot agent is available has its own set of connected sources (tools) that determine what Autopilot can access and do.

Here are the main capabilities of the chat experience:

• Context retention—remembers previous inputs and decisions.
• Prompt refinement—lets you edit your prompt in real-time to improve outcomes without starting from scratch.
• Conversational experience—uses natural language to ask questions and make changes.
• Inline suggestions—offers recommendations directly in the chat, specific to your current context.
• Unified workspace—keeps all interactions, iterations, and results in one place.
• Workflow continuity—maintains the flow across steps and stages.
• Interactive debugging—helps you identify and fix issues by asking follow-up questions or clarifying intent.
• Code and logic awareness—understands the structure of your automation or code and provides relevant guidance.

Interaction modes

In the user interface, the Edit dropdown in the chat input area allows you to switch between two interaction modes:

• Edit — Requires user confirmation to perform change actions.
• Agent — (Experimental) Performs actions automatically without requiring user confirmation.

Available models

Autopilot agent supports multiple LLMs configured through AI Trust Layer. Some models are designated as "Agent" models, optimized for tool-calling and multi-step workflows. The model selection is available from the chat panel.

The choice of LLM influences performance, reasoning depth, and speed.

• Gemini 2.5 Flash—ideal for fast, general tasks without deep reasoning needs.
• Gemini 2.5 Flash (Agent)—combines speed with agentic capabilities for moderately complex, multi-step tasks.
• Gemini 2.5 Pro—best for complex automation building, reviewing, and debugging, offering deeper reasoning.
• Gemini 2.5 Pro (Agent)—the top choice for highly complex, agent-driven tasks requiring thoroughness and advanced planning.
• GPT-5 Mini - Experimental—for simple tasks where quick response time is paramount; an experimental model.
• GPT-5 Mini - Experimental (Agent)—experimental agentic model for quicker, structured outputs on simpler tasks.
• GPT-5 - Experimental—for general to complex tasks, leveraging cutting-edge, experimental LLM features.
• GPT-5 - Experimental (Agent)—experimental agentic model for complex, agent-driven tasks, using the newest models.

The (Agent) suffix indicates that the model is specifically configured to leverage agentic capabilities, such as:

• breaking down complex problems.
• selecting and using appropriate tools (Web Search, RPA Workflow Assistant).
• executing multi-step processes.
• monitoring progress and adapting its strategy.

If your task involves multiple steps, requires using tools, or needs deeper reasoning beyond a single response, opting for an (Agent) variant is generally recommended.

You can select the LLM from the Autopilot chat screen.

Note:

• To use these LLMs, check you have the required policies in AI Trust Layer.
• Specifically for the documentation AI functionality, Autopilot uses Azure OpenAI.

Kimi K2.6 support

Depending on your Studio version, the Kimi K2.6 model appears under a different display name:

Product and version	Display name
Studio Desktop 2026.0.x (STS releases)	Helix
Studio Desktop 2025.10.x (LTS releases)	Virtuoso
Studio Web in Automation Cloud	Virtuoso
Studio Web in Automation Suite 2.2510.x	Virtuoso

No action is needed on your side. Starting with the 2026.10 releases, these interim names go away: Autopilot will display the model under its actual name, Kimi K2.6.

The AI Trust Layer policy

To use the AI capabilities provided by Autopilot, make sure you have an AI Trust Layer policy active and deployed to your tenant.

Learn how to:

• Create a policy
• Configure the policy for AI Trust Layer
• Deploy the policy to the tenant where you want use Autopilot

Disclaimer about Autopilot responses

Autopilot is designed to assist you across different UiPath products, and provides AI-powered support for a variety of tasks, such as building and editing automations, understanding complex workflows, or retrieving relevant information from documentation or other sources. While it can be highly effective, it may occasionally generate responses that are inaccurate, misleading, or incomplete. This is a known limitation of large language models and is sometimes referred to as "hallucination."

For example:

• Autopilot may suggest steps that do not apply to your current workflow.
• It may provide explanations that sound correct but are factually incorrect or based on outdated logic.
• When working with external information, such as documentation or web search, it may overlook important details or misinterpret source content.

Always review and validate its responses, especially when dealing with production workflows, data handling, or error-prone logic:

• Double-check suggestions before applying them, particularly in critical or high-risk workflows.
• Use the thumbs down button in the chat interface to flag unhelpful or incorrect responses.
• Review documentation links or search citations if provided. These often contain additional context not included in the summary.

Autopilot is a powerful assistant which works best as a collaborator, not a decision-maker. Your input and judgment remain essential for building safe, accurate, and effective automations.