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

Starting December 2, 2026, the AI Trust Layer governs low-code and coded agents as one product. The agents settings in your AI Trust Layer policies and the LLM configurations defined for low-code agents apply to coded agents too. The change applies to the LLM calls that coded agents make through the AI Trust Layer. Existing coded agents need no code changes or redeployment.

On that date, the coded agents settings will be hidden from AI Trust Layer policies, LLM configurations, and Model hub, and will no longer be enforced.

Wherever the two sets of settings differ, coded agents change behavior:

Before December 2, 2026, check that the agents settings in each AI Trust Layer policy are how you want coded agents to behave: compare the product toggles, guardrails, and allowed models with the coded agents settings. To keep coded agents on their current settings:

For example, if your coded agents settings allow only GPT-5 and your agents settings allow all models, coded agents can use all models after the change. To keep the restriction, create two AI Trust Layer policies and deploy them by group:

Use the same approach for differences in the product toggles and guardrails.

If you take no action, your agents settings apply to coded agents as they are. You can update your policies and their deployments at any time, including after December 2. After the change takes effect, check that your coded agents behave as expected. Pricing, licensing, and stock keeping units (SKUs) do not change.

For details, refer to Agents governance, Settings for AI Trust Layer policies, Configuring LLMs, and the Deprecation timeline.

You can now keep validating your agents after they reach production, not only while you build them. Runtime evaluations score a deployed agent against datasets of inputs and expected outputs, on demand or on a recurring schedule, so regressions, accuracy issues, and unexpected behavior surface without a manual testing cycle.

Because the evaluations run against the deployed agent, the results reflect what is in production rather than a design-time copy. Developers, solution architects, and agent quality and governance teams get ongoing visibility into how reliable each agent stays after deployment. Production runs can also become test cases, so your datasets grow with the cases the agent handles.

For details, refer to Managing runtime evaluations (Preview).

You can now add a dedicated Teams app to a standard Teams channel. Anyone in the channel can mention the agent in a thread, and it replies in that same thread, taking the recent messages in the thread into account.

Channel users don't sign in to UiPath. An organization administrator registers the UiPath-provided Conversational Agents external client once for the organization, and the agent runs under the identity of an unattended robot account assigned to its Orchestrator folder. Private and shared channels and file sharing in channels are not supported.

For details, refer to Dedicated Microsoft Teams App.

Correction, October 28, 2026: This feature was released on October 14, 2026, but was not included in the release notes at the time.

Agent review replaces the Agent score. Ask Autopilot to review your agent, or select the Review agent badge. The review finds critical issues, warnings, and improvements, and grades the agent from A to F based on the number and severity of the issues. Autopilot can fix the issues and apply the improvements in the same conversation.

For details, refer to Agent review.

Your agents can now be protected by your organization's own subscription to a supported safety vendor, not only by the UiPath-managed services. Once an administrator brings a subscription into the AI Trust Layer, its guardrails become available to you alongside the UiPath ones, each marked as BYO and showing which provider evaluates it. Azure AI Language, Azure Content Safety, and Noma are the vendors supported during preview.

You choose and configure them as you do the out-of-the-box guardrails, with the same scopes and actions, and every detection appears in the agent trace together with the provider that produced it.

For details, refer to Configuring guardrails and Out-of-the-box guardrails.