2026.21.0
Cognigy
Released October 12, 2026Features
New Models by Mistral AI
Added support for Mistral AI’sministral-14b-latest, ministral-8b-latest, and ministral-3b-latest. These models are compact, offer a strong performance-to-cost ratio, and support several languages, including English, French, Spanish, German, Japanese, and Arabic. Use ministral-3b-latest for fast, lightweight tasks, and ministral-14b-latest when you need higher accuracy at a low cost.
Improvements
- Added support for the A2A Agent Node as a child Node of the LLM Prompt Node
- Added
API Keyto the Performed By column and filter of the Audit Events page for changes made through the API
Bug Fixes
- Fixed the issue where the Progress Message field in the Call A2A Agent, Call MCP Tool, and Execute Workflow Tool Nodes had no effect except on A2A Server Endpoints. Now, the field is removed from these Nodes, and the Call MCP Tool and Execute Workflow Tool Nodes no longer send status updates to A2A clients
- Fixed the issue where sending messages in conversations already handed over to a human agent failed after restoring a Snapshot, because the restore deleted the handover provider and created it again with a new ID. Now, the handover provider isn’t deleted but updated, and it keeps the same ID
- Fixed the issue where OData Analytics requests were rejected with
429 Too Many Requestsduring normal reporting use - Fixed the issue where follow-up requests after an AI Agent tool call to Google Gen AI models, such as
gemini-3.5-flash, contained the tool-call ID instead of the tool name - Fixed the issue where a streamed OpenAI or Azure OpenAI Responses API answer that consisted of two text messages separated by a tool call repeated the first sentence and dropped the first word of the second message
- Fixed the issue where calls to the Gemini 2.5 model family failed with
Value is not a valid number: NaNwhen the model returned an empty answer - Fixed the issue where API requests were slower because the permissions cache wasn’t used after a version mismatch between services, so permissions were recomputed on every request
- Fixed the issue where
service-apiloggedA cacheClient.get call has erroredand skipped the cache because the user cache and the permissions cache used the same Redis key
Cognigy Voice Gateway
Released October 12, 2026Bug Fixes
- Fixed the issue where the Click To Call widget intermittently failed to render when initialized on the same page as Webchat
- Fixed the issue where inbound SIPREC calls ended right after being answered when the Flow responded too early
Cognigy Agent Copilot
Released October 12, 2026Improvements
- Added the End Session On Handover Close toggle and the Session TTL On Handover Close field to the Copilot section of the Endpoint settings. By default, the Agent Copilot session ends when the human agent resolves the conversation. When you deactivate the toggle, the session stays open for the number of minutes set in the field (default 120), so the human agent can keep sending data to the Agent Copilot Flow from the Agent Copilot workspace
- Added the Enable Branding toggle to the Copilot section of the Endpoint settings. This toggle controls whether the
Powered by Cognigy.AIbranding appears at the bottom of the Agent Copilot workspace
Cognigy Live Agent
Released October 12, 2026Improvements
- Improved service stability and performance