September 2026 release notes for Agents in Automation Cloud and Test Cloud.

Agents can now use Data Fabric entities directly as a structured context, letting them query and reason over live or near-real-time, structured system of record without relying on static data snapshots. This lets users query this data in natural language, while existing enterprise access rules and permissions are enforced.

For details, refer to Contexts.

Agents can now be built on the advanced harness, an alternative to the standard reason-and-act loop for long-running, document-heavy work. An advanced agent keeps its intermediate work in a file workspace instead of in the prompt, delegates slices of work to sub-agents and calls its tools programmatically from a sandboxed JavaScript interpreter.

During public preview, advanced agent consumption is 50% off the standard rate.

For details, refer to Advanced agents (Preview).

Building UiPath agents with a coding agent is now generally available. You can build both low-code and coded agents entirely from a coding agent (Claude Code, Codex, Cursor, and others) and terminal.

The uipath-agents skill covers the full lifecycle for both agent types — scaffolding a project, editing prompts and input and output schemas, adding tools, running evaluation sets, and packaging and deploying to Orchestrator. The skill is available to install via uip skills install.

For details, refer to the following pages:

Chat agents can now be built as a Maestro Flow, alongside the existing single-prompt Agent Builder path. Modeling the chat experience in Flow unlocks capabilities like intent-based routing, hand-off between specialist agents, deterministic replies, and parallel branches that run while the conversation continues. Both build paths surface to end-users as the same chattable Conversational Agent, meaning you can switch to Flow without disrupting how users reach your conversational agents.

For details, refer to Chat.

Voice agents can handle real-time phone calls — answering inbound calls and placing outbound ones. Audio streams straight to a real-time speech model, so callers can interrupt the agent mid-sentence, and tools run during the call.

You can hand off between specialist voice agents without dropping the call, feed the call transcript or a structured call summary into downstream nodes, and let the agent press digits to navigate an external phone tree. Numbers connect over a SIP trunk or your own Twilio account.

For details, refer to Voice.

The agent design policies in the Automation Ops governance policy for Studio Web are now generally available. The policies let you enforce the following agent configuration requirements before an agent is published:

For details, refer to Settings for Studio Web policies.

A new built-in guardrail, LLM as Judge, is now available in preview. LLM as Judge uses an LLM to evaluate agent prompts, responses, or LLM calls against custom instructions that you write in plain natural language, closing the gap left by other out-of-the-box guardrails, which only detect a fixed, predefined category through a managed classification service.

Each check the guardrail performs makes a real LLM call, so the judge model consumes Agent Units or Platform Units on top of the agent's own usage, tracked separately. This applies wherever the guardrail runs, including agent evaluation runs with the guardrail enabled. For details, see Out-of-the-box guardrails and Licensing.