Release v3.9.0 brings browser-based voice calling to AI Employees, expands outbound SMS campaigns with CSV uploads, and re-engineers the Assessor for scalable, multi-channel testing. Alongside these headline features, this release delivers major stability improvements to multilingual support, cross-channel conversations, and action execution reliability.
WebRTC browser-based voice calling
AI Employees can now receive voice calls directly from web browsers using WebRTC technology. The new newo_voice_connector_web connector enables a lightweight JavaScript widget that partners can embed on any website, giving visitors a one-click call button — no phone number or app download required. The widget supports audio and video modes, custom theming, and programmatic control via JavaScript API. Inside the platform, agents can also be tested immediately after creation using the in-app calling feature.
Bulk outbound SMS via CSV in Workflow Builder
Partners can now launch automated outbound SMS campaigns by uploading CSV files through the Workflow Builder (Widgets) interface — matching the existing functionality for outbound calls. Create an item in the Outbound SMS CSV files column, paste structured CSV data with phone numbers and conversation reasons, and the AI Employee handles personalized greetings, two-way SMS conversations, session management, and automated result tracking. This feature is powered by the new integration between the Outbound module and Workflow Builder.
Assessor client-server architecture
The Assessor module has been re-engineered with a client-server architecture, migrating to a new assessor_server for better coordination of Twilio resources and assessment queues. This enables high-scale parallel testing, reducing validation time as agent portfolios grow. The architecture supports both voice and chat evaluation workflows.
Assessor web chat testing
The new AssessorChatFlow enables end-to-end automated evaluation of text-based web chat agents. Partners can now run structured test scenarios against chat agents — verifying responses, conversation flows, and business logic — alongside existing voice testing capabilities. Combined with broadcast level tracking, assessments now capture comprehensive system events and agent thought processes.
Language switching stability
Multilingual support receives a significant reliability upgrade with two new rules: the Anchoring Rule (2-turn persistence) locks the detected language for at least two conversation turns, and the Stability Rule prevents accidental language switches caused by STT noise, single-word inputs, or incidental utterances. These rules have been applied across all industry templates.
Wait for action execution
A critical reasoning improvement prevents the agent from "hallucinating" success and proceeding to subsequent conversation steps before actually receiving results from background tasks. For example, the agent will now wait for an availability check to return before asking the customer for their name — eliminating premature responses that confused callers.
Bug fixes
Working hours accuracy: Fixed a bug where the agent incorrectly reported being "Closed" when it was "Open" because the minute value was being replaced by the hour value in the internal time representation.
Multi-channel context persistence: Fixed a bug where the agent would "forget" conversation details when switching from a phone call to SMS within the same session (e.g., repeating greetings instead of continuing the conversation).
Webhook query handling: Fixed an issue where custom tools using GET webhooks would fail to send query parameters correctly.
Multilocation first-time setup: Fixed a failure occurring when new accounts enabled the Multilocation feature for the first time.
Library updates: Fixed an issue in the Workflow Builder where adding a new item to the library would not update the agent's available instructions immediately.
Prompt syntax: Corrected an incorrect closing tag
</ActionStates>(should be</ActionsStates>) in the prompt builder.International character support: Resolved display issues with special characters and non-English text by standardizing JSON serialization settings across all agent modules.
SMS & email placeholders: Fixed an issue where automated messages contained messy empty fields when information was missing; the system now cleanly removes both the empty placeholder and its label.
ManyChat reply type: Fixed an error where user IDs were passed as integers instead of strings, preventing the agent from replying to messages. (Previously hotfixed.)
ManyChat TikTok flow stability: Fixed a crash during integration setup when the ManyChat API returned empty or unexpected data for TikTok flows.
Outbound session orchestration: Introduced
outbound_session_globalto ensure reliable session tracking and prevent duplicate outbound attempts.XSS sanitization: Added security sanitization for data retrieved from external websites to prevent untrusted content from entering the RAG context. (Previously hotfixed.)
Restaurant scenario typo: Corrected the cancellation scenario reference from "Cancellation" to "Canceling Existing Reservation."
JSON serialization standards: Applied standardized JSON serialization across all modules.