Skip to main content

Release 5.0 2026.09.08.1

· 7 min read

New Features 🚀​

Agent To-Dos [Preview]​

Agents can now break a request down into a task list and work through it step by step, reporting progress as each item is completed. This makes long, multi-step executions easier to follow.

An agent preview showing a Tasks panel with four completed items, each marked Done, above the generated guide

This feature is currently in preview.

User Choices [Preview]​

Agents can now pause and ask the user to pick from a set of options before continuing. The choices are presented as selectable buttons in the conversation, enabling richer interactions and ease of use. Available in the Chat Widget and the Serenity* AI Chat. See Human-in-the-loop for integration guides.

A chat asking Choose an option, with the question What kind of bike are you looking for? and the options Road bike, Mountain bike and Hybrid / city bike

Two answered question groups in the conversation, each collapsed into a summary of the questions and the answers given

This feature is currently in preview.

Agent Onboarding​

The onboarding flow now recommends the agents that best fit your organisation's profile, so a new tenant can start from a relevant set of agents rather than from an empty workspace. Pick the ones you want and they are created in your tenant, ready to use.

Agents that come with supporting functionality, such as workflows or supporting agents, can now be configured through a step-by-step flow when you get them from the Store. Making it faster to set up and ensures no required configuration is left out.

The Agents made for you onboarding step, suggesting agents such as Coding Expert and Customer Support Assistant based on your company

The Finish setting up your agent screen for SEO Analysis, showing the configuration it still needs

Inference Endpoints for All Models​

OpenAI-compatible inference endpoints are now available for every model in the catalogue, so any model can be called directly from your own code with the same interface.

The Inference playground with Gemini 3.8 Flash selected and the Responses API chosen, beside a successful execution result

OpenAI Privacy Filter​

The OpenAI Privacy Filter is now fully available as an AI Service. Instead of hosting your own implementation, you can call our hosted Privacy Filter directly to detect and redact personal data such as names, addresses, emails and phone numbers, either one text at a time or in batches.

The Privacy Filter playground calling the batch method, with a request redacting names and addresses and the redacted response beside it

Enhancements ✨​

Request Access to Connectors​

When a connector is not enabled for your company, users can now request its activation from within the platform. Administrators review and approve the request, which complements the connector policies defined in the Compliance Center.

The Available connectors list showing Notion MCP marked Access requested, with the name of the user who requested it

Comments on Message Feedback​

Feedback on an agent message can now include a comment explaining the rating, instead of only a positive or negative signal. Available in the Chat Widget, the Serenity* AI Chat and the feedback endpoint.

An agent message rated negatively, with an optional What went wrong? comment box below it

Reasoning Effort and Output Format in Proxy Agents​

Proxy agents now accept the reasoning effort and the response format parameters, keeping compatibility with the OpenAI request schema and giving you control over structured outputs and how much the model reasons.

Cost Savings in Billing​

The billing screens now report the savings obtained through your plan, the credits billed to your Serenity* account and the estimated usage when the Bring Your Own Key feature is in use.

The Cost Analysis panel showing, for each period, the cost before the plan discount struck through, the final cost and the amount saved with your plan

Time to First Content​

Executions now report the time to first content, which measures how long a user waits before seeing the first part of a response. This complements total execution time when diagnosing perceived latency.

The executions grid with the new Time to first content column highlighted beside Time to first token

Active and Inactive Agent Versions​

Agent versions can now be activated and deactivated, so a version can be taken out of service and avoid unwanted use of stale versions.

A confirmation dialog warning that a deactivated version will no longer be executable through the API and that open conversations pinned to it will stop working

Downloadable Knowledge Files​

Knowledge files can now be marked as available for download, which makes them accessible from citations so logged-in users or clients can open the cited file. Downloading files from citations is supported in the Chat Widget and the Serenity* AI Chat.

Files marked as available for download are also added automatically to the workbench session, so the agent can work with them directly.

Configure download availability through the knowledge metadata in the Agent Design Studio.

The knowledge file metadata dialog, with the Available for download toggle enabled below the description and tags

A citation in the chat widget showing the cited PDF with a download button beside its name

Bug Fixes & Minor Improvements 🛠️​

  • Knowledge sources can now be used as system parameters with {{system.knowledge}}.
  • Improved error handling in the AI steps of the tenant onboarding.
  • Improved error handling in the dynamic content generation of HTTP Request skills.
  • Fixed the client-side colour configuration in the chat widget.
  • Fixed an error that prevented quotas from being set on tenants without an assigned plan.
  • Fixed an error in the discounts management screen that allowed 0% discounts to be added.
  • Permission selectors no longer list system users such as api-key-user.
  • Improved the pending changes indicator when sharing a version.
  • Fixed an error that prevented a file from being uploaded to the dropzone if the same file had been uploaded, removed and uploaded again.
  • Fixed the count of agents using a connection in My Connections.
  • Added an endpoint to query the tenant's monthly usage.
  • Fixed an error in Long Term Memory that generated two identical slugs.
  • Improved texts, captions and calls to action across several areas of the platform.
  • Improved the tooltips in the Agent and Cost Manager dashboards.
  • Internal plugins are now hidden from the skills catalogue API.
  • Fixed a null reference exception returned when executing the /conversation endpoint with an Activity Agent.
  • Fixed an issue in the Inference Playground code editor that made the page unresponsive when streaming very long responses.
  • Improved ElevenLabs vendor billing by using the Credit dimension.
  • Fixed an error in WhatsApp when sending a PDF with no text in the user's message.
  • Fixed an error in the datasets skill where a value was treated as required when it was not.
  • Fixed an error in the manual renewal of a subscription that did not regenerate the tokens included in the plan.
  • AI Service executions and workflow instances are now cleaned up based on the tenant's subscription plan settings.
  • Compatible web search models now execute as provider server tools.