2026-09-16 – 2026-09-22
about 2 hours ago by ReadMe API
🚀 New Features
- Freshdesk is now available as a knowledge source. A Freshdesk helpdesk can be connected to an agent so its tickets become agent knowledge.
- Connection is made with the Freshdesk subdomain and an API key, entered on the Create Agent page under Customer Support.
- Tickets and their replies are imported. Internal agent notes are never included.
- An anonymization option removes personal details from imported tickets.
- Available on all plans, with auto-sync available on Enterprise.
- Freshdesk's own per-minute API rate limits apply, which depend on the helpdesk's own plan.
- LinkedIn is now available as a knowledge source. Public posts from any personal profile or company page can be turned into agent knowledge.
- The profile or company page URL is pasted on the Create Agent page or the Sources page.
- Posts and reposts that carry a comment are synced. Comments left underneath a post are not.
- Auto-sync is available on all plans, and behaves the same way it does for sitemaps.
- Stop Response has been added to Enterprise Agents. A response can now be stopped while it is being produced.
- The send button becomes a stop button for the duration of the turn, and returns to send once the final answer starts being written.
- The agent finishes the task it is already working on and then stops. A turn can also be stopped before streaming has started.
- Query credits are counted for every task that ran, including the one in progress, and not for tasks that never began.
- The agent can resume the work if it is asked to continue.
- Works on every deployment - the dashboard, the embed widget and shared links - and through the Tasks API.
- Available on agents running Plan & Act only.
- Citations are now fully supported on Multi-Agents.
- A Citations tab has been added to the Multi-Agent's Personalize page, carrying the same Citation Settings, Citation Types and PDF Citations sections a regular agent has.
- Those settings override the child agent's own citation settings for the whole conversation, so the presentation stays consistent as the Multi-Agent moves between children. Advanced Settings continue to be read from the child.
- The PDF Viewer is supported.
- Citations also appear on Multi-Agent messages sent before this shipped, where the message itself carried them.
🛠 Improvements
- Zapier now connects through a CustomGPT.ai authorization screen instead of a manually created API key. The connection grants access to selected agents only, and managed connections are listed under My Profile > Connections.
- Enterprise Agents plan and replan faster, because the task planner and evaluator now read only the instructions for the tools that are actually enabled.
- Enterprise Agents are better at pulling the relevant part out of a long document, because the agent can now point at the section it wants and find it directly in the file.
- An agent's chat now loads on its own background colour, or on a preview of its background image, instead of a generic grey screen. This covers the dashboard, embeds and shared links.
- On Enterprise custom plans, the per-user API key limit can be raised above the standard 20, to a maximum of 500. A user who belongs to several teams receives the highest limit among them.
- In-agent drag and drop now accepts a file dropped anywhere on the agent surface, including the conversation history and Files sidebars.
- In the Enterprise Agents composer, the folder icon is hidden when the file panel has nothing to show.
🔌 API & MCP
- Added
POST /api/v1/projects/{projectId}/tasks/{taskId}/stop, which stops a running Enterprise Agents task. - The MCP server now hides the tools that the connected agent's chat flow rejects, so a calling model can no longer pick a tool that is guaranteed to fail. Task tools are hidden on agents without Plan & Act, and
query_customgpt_knowledge_baseis hidden on agents with it. - Customer Intelligence records retrieved over MCP now carry
created_at. The field was previously dropped before it reached the client. - Multi-Agents can now be renamed and deleted through
POST /api/v1/projects/{projectId}andDELETE /api/v1/projects/{projectId}. Both previously returned403 This action is unauthorized. POST /api/v1/projects/{id}/actions/customnow returns a422namingmcp_urlwhen that field is missing, instead of a500.- Requests carrying text that is not valid UTF-8 now return a
400naming the offending field, instead of a500.
🐞 Bug Fixes
- Fixed the issue where an Enterprise Agents answer listed more citations in the citation drawer than it showed inline in the response.
- Fixed the issue where an agent that called two MCP tools at once lost the result of one of them, and answered without that data.
- Fixed the issue where an External API connection returned a
401on a Multi-Agent, although the same connection worked on a standalone agent. - Fixed the issue where the Publish modal showed a public agent as private when it was opened from the Actions tabs, the External API screens or the Event Log.
- Fixed the issue where messages produced by a scheduled task appeared in analytics with a blank location and an unknown browser, and were grouped with unidentified visitors on the Location card. They are now labelled Scheduled Task across analytics and Customer Intelligence, and exports write
scheduled_task. - Fixed the issue where the disconnected-integration warning banner overlapped the page beneath it when its message ran to more than one line.
- Fixed the issue where the file upload icon in the chat preview appeared only on the Actions page, did not react to the Plan & Act toggle without a reload, and could render a duplicate send button.
- Fixed the issue where the conversation menu on the Ask page mixed languages, because one of its items was never translated.
