July 2026

July 2, 2026

Conversational agents: new runtime, features and changes

Low-code conversational agents now run on unified runtime, the same execution foundation that low-code autonomous and coded agents already use. This brings new features and capabilities, as well as some important behavioral changes.

Nouvelles fonctionnalités et améliorations

• Inputs — You can define inputs on the conversational agent and template them into the system prompt and tool-call arguments. You can set and update inputs throughout the conversation, allowing the agent to receive additional context variables beyond the conversation history. For details, see Conversational agent inputs.
• Tool-call confirmations — You can optionally configure tools to pause before execution, allowing the user to review, approve, modify, or reject the proposed input parameters before the tool runs. For details, see Tool-call confirmations.
• Tracing — The runtime now represents each conversational agent execution as a single Orchestrator job per agent response. This brings greater visibility into the inputs, lifecycle, and tracing for each conversational agent execution. For details, see Conversational agent observability.
• Improved citations — Greater consistency and stability for citation generation when the agent references websites and Context Grounding indexes. We also fixed the citation previewer for when the iFrame is embedded within a UiPath App.

Behavior changes

• Runs as the Cloud user's identity — When an authenticated user in your Cloud organization invokes a conversational agent during a conversation, the agent now runs under that user's identity. For details, see Execution identity.
  • To ensure tool executions match the previous runtime's behavior, users must have the required access for certain tool resources (such as Context Grounding indexes).
  • The host folder will no longer always require an unattended robot account for execution, and you should always assign a serverless machine template instead. For details, see Orchestrator folder requirements.
• Sequential tool processing — When a conversational agent calls multiple tools in parallel, the runtime now processes those tool calls one at a time instead of concurrently. This ensures more stable handling of guardrails and potential errors during tool execution.