Autopilot for Agents

Autopilot for Agents (Preview) is a conversational AI assistant embedded in the Agent Builder in Studio Web. It provides a persistent, context-aware workspace where you can build, configure, and refine agents using natural language, without switching between tools or manually editing each component.

Autopilot reads your current agent definition and available platform resources each time a session opens. When you describe a change, Autopilot proposes targeted edits to your agent—to its prompts, tools, schema, or context sources, which you review before applying.

The chat interface and general interaction model are shared across all UiPath products that surface Autopilot. For a full reference of common interface elements, see Autopilot chat user interface. This page covers the capabilities and tools specific to the Agent Builder context.

Schlüsselfunktionen

Autopilot for Agents provides the following capabilities:

• Context retention — remembers previous inputs and decisions within a session.
• Prompt refinement — lets you adjust your prompt in real time without starting over.
• Inline suggestions — proposes changes specific to your current agent state.
• Workflow continuity — maintains context across iterative changes and refinements within a session.
• Interactive debugging — helps identify and clarify issues through follow-up questions.
• Code and logic awareness — understands the structure of your agent and provides guidance relevant to its current state.

How Autopilot for Agents works

When you open the Autopilot panel using the widget on the right-side toolbar of the Agent Builder, Autopilot fetches the current state of your project, including:

• All agent runs and evaluation set runs
• The latest health score
• Available resources: RPA processes, API workflows, agent processes, and Context Grounding context sources

This context is included with every message you send. The Agents backend initiates a streaming request to the selected large language model (LLM), combining your message with the agent definition, available resources, and a system prompt template. The response streams back to the chat panel.

Autopilot surfaces suggested changes as previews. In the default interaction mode, you review and accept each change before it is applied to your agent.

Interaction modes

The chat input area includes an Edit control that sets how Autopilot applies changes to your agent:

• Edit (default) — Autopilot proposes changes as previews. Each suggestion requires your confirmation before being applied. This mode is recommended for all agent design and production workflows.
• Agent (experimental) — Autopilot applies changes automatically without requesting confirmation. This mode is intended for exploratory or rapid iteration workflows.

Wichtig:

Agent mode is experimental. Use it with caution when working with production agents.

The chat interface

The Autopilot panel is accessed from the right-side toolbar of the Agent Builder in Studio Web. The panel provides:

• A chat input field for entering prompts and questions.
• Dynamic prompt suggestions that update alongside your agent's current state.
• An LLM selector for choosing the language model for the session.
• File upload support for documents, images, code files, and data formats, up to 15 MB per file.
• New chat, Settings, and Chat history controls in the panel header.

Chatverlauf

The Chat History panel provides a searchable list of conversations from the last 30 days. Selecting an entry restores its full context, letting you continue where you left off or reference previous suggestions.

Einstellungen

The Settings panel controls how Autopilot behaves in the current Agent Builder session.

Personalization:

• Enable Prompt Suggestions — on by default; displays contextual suggestions after each response.
• Show Context Change Messages (Experimental) — displays messages when Autopilot updates its understanding of the agent context.
• Custom Instructions — a free-text field, up to 1,000 characters, for specifying terminology preferences, desired detail level, or constraints on Autopilot responses.

MCP Servers:

Connects Autopilot to Model Context Protocol (MCP) servers to provide additional structured context. Only MCP servers created in Orchestrator are supported.

Tools:

Displays the Framework Tools available in the current session. Each tool can be individually enabled or disabled. For details, see Autopilot for Agents tools.

Save applies your current settings. Reset to Defaults reverts all settings to their original state. You can also Download conversation to export the current session as a JSON file that includes organization, tenant, and conversation settings.

Antwortgenauigkeit

Autopilot for Agents provides AI-powered support for agent design tasks such as configuring prompts, adding tools, and refining schemas. 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.

Zum Beispiel:

• Autopilot may suggest tool configurations that do not match your available resources.
• It may provide explanations that are factually incorrect or based on outdated agent logic.
• When working with external resources, it may overlook important details or misinterpret source content.

Review all suggestions before applying them, especially for production agents:

• Use the thumbs down button to flag unhelpful or incorrect responses.
• Prefer the default Edit mode over Agent mode for changes to critical agent logic.

Autopilot is most effective as a collaborator. Your judgment remains essential for building accurate and reliable agents.

Ähnliche Themen

• Autopilot for Agents tools — reference for Framework Tools available in the Autopilot panel
• About Autopilot chat — overview of the Autopilot chat framework across all UiPath products
• Autopilot chat user interface — full reference for the chat interface shared across all products