Kore AI Release Notes

Minor Release

This update includes new features and enhancements summarized below.

Multi-Agent Orchestration

Integration with Microsoft Agent 365

Agents can now be integrated with Microsoft Agent 365, enabling organizations to export and register agentic applications directly into their Microsoft 365 tenant. This allows agents to be used in Microsoft 365 Copilot and Teams, with centralized governance and telemetry through Microsoft Agent 365. Administrators can manage identity, lifecycle, and observability using standard Microsoft Entra and Admin Center controls while continuing to build and run agents in Agent Platform.

Import and Export APIs for App SetUp

Public Import and Export APIs are now available for programmatic management of app lifecycles. Developers can import complete app setups with scope-based validation and export specific app versions as .zip files, supporting automation and CI/CD workflows.

Learn more →

AI Engineering Tools

Expanded Model Support

The platform now supports additional AI models:

• OpenAI: gpt-5.5-2026-04-23
• Anthropic: claude-opus-4-7

No-Code & Pro-Code Tools

Role-Based Access for Workflow Tools in Agentic Apps

Workflow tools within an Agentic app no longer require separate access management. They inherit access from the user’s app-level permissions. If a user has access to the app, their role determines whether they can access the associated workflow tools. Each app role includes a dedicated permission that controls this access. This eliminates the need for individual tool sharing or tool-level permission management, ensuring that access is consistent, centralized, and directly tied to the app.

Learn more →

Original source

This update includes enhancements and bug fixes. The key enhancements included in this release are summarized below.

Integration Enhancement

Custom Data Mapping for Salesforce Chat Messaging

Agent AI now supports JSON-based mapping to pass Salesforce chat messaging object data — such as agent name and case details — to the widget during load. Admins can map each key to a payload field name and each value to a Salesforce Case object field on the Agent AI Configuration page.

Widget Enhancement

Auto-Trigger Next Best Action for Web-to-Case Records

The Agent AI widget automatically triggers the first Next Best Action (NBA) when a Web-to-Case record is opened, using configured case and customer fields as the input query to the Agentic app. Subsequent NBAs trigger based on incoming customer emails. You can select Override at any time to fetch an NBA manually. Authorized Salesforce developers can update NBA query fields without code changes.

UI Enhancement

Agentic Response Analytics on Agent AI Dashboard

The Agent AI dashboard now includes Overview and Feedback analytics covering metrics for Agentic Copilot responses and suggestions, as well as overall widget performance. Admins and supervisors can view feedback comments and reasons, track copy, send, and override AI suggestion usage. They can also monitor performance metrics — such as end-of-conversation generation time and average AI suggestion generation time — to identify areas for improvement. Learn more→

Original source

This update includes enhancements and bug fixes. The key enhancements included in this release are summarized below.

Console

Notes Translation

Notes in the Console and the Conversations Dashboard now support translation based on the user’s selected language, with translated-only and dual-view modes available. No additional configuration is required.

Configuration

Restrict Translation Languages to User Assignment

A new Limit Translation Languages to User's Assigned Languages setting can restrict the translation dropdown to show only the logged-in user’s assigned languages across Agent Console, Monitor, and Interactions.

Disposition Capture During Transfers

A new Capture Disposition During Transfer setting displays a dropdown during agent or supervisor transfers, allowing optional selection of one or more dispositions. Selected dispositions are stored with the interaction and remain available for reporting, while transfers continue without interruption if skipped.

Analytics

Interaction Details Report Updates

A new Include In-Progress Conversations option displays active conversations with their current status and wait time. Agent Custom ID, a new Support field, and a configurable date format have also been added.

Open Wallboards in Separate Window

A new Launch Wallboard in New Window option is available in both the Configuration and the Wallboards views, opening the selected wallboard in a separate browser window for independent viewing.

Original source

This update includes new features and enhancements summarized below.

Multi-Agent Orchestration

New Input Processor

A new Input Processor capability enables preprocessing of user input before the app processes it, allowing centralized handling of authentication, input transformation, and context enrichment. It works across all orchestration patterns, reduces duplication across agent pre-processors, and ensures consistent, application-wide behavior.

Learn more →

Preprocessors for Events

Pre-processors for system events — welcome, agent handoff, and end of conversation — are now supported, enabling developers to execute custom logic before event actions are triggered. These scripts can access and enrich context, perform transformations, or call external systems before the event is processed.

Learn more →

AI for Service Logs Integration

Integration of logs from AI for Service helps centralize interaction data, providing access to detailed model telemetry, including token usage and request/response payloads. This enables improved monitoring, analytics, and billing accuracy. This feature is currently in preview and can be enabled upon request.

Agent Protocol Enhancement - Get Sessions API

A new Get Session API is available to fetch details for a specific session using session_id. The API supports two include parameters: traces and observations, to return detailed execution data, including traces and nested observations. This enables deeper visibility into session behavior, model interactions, and telemetry for debugging and analysis.

Learn more →

AI Engineering Tools

Expanded Model Support

The platform now supports three additional Google Gemini models:

gemini-3.1-pro-preview

gemini-3.1-flash-lite-preview

gemini-3.1-pro-preview-customtools

Learn more →

Original source

Integration Enhancements

Internal Transfer Event Support in Salesforce Omni Chat

The Agent AI widget now supports internal transfer events in Salesforce Omni Chat integration, recording NA entry and exit events, and agent details during chat sessions.

Hybrid API Support for Salesforce Chat Data Retrieval

The Salesforce Agent AI widget now fetches chat landing summary data using the Conversation Toolkit API, with automatic fallback to the Connect REST API when message data is unavailable. This ensures a reliable conversation summary is always available without latency or data gaps.

Real-Time Call Transcription Streaming via WebSocket

Agent AI now supports WebSocket-based streaming of real-time call transcription to the Agent AI backend, delivering low-latency transcription data during active calls and enabling real-time assistance and analytics without webhook-related delays.

Widget Enhancement

Conversation Metadata for Troubleshooting and Root-Cause Analysis

Agents can now copy conversation metadata from the Agent AI widget under Settings > Help & Support, helping support teams troubleshoot issues faster and perform root-cause analysis.

UI Enhancements

Bot-User Conversation Support in End-of-Conversation Summary

Admins can now include bot-user conversations in the End of Conversation Summary. Two summary types are available Full Summary (bot-user and agent-user conversations) or Agent-User Summary Only. By default, only the agent-user conversation is included. The bot-user conversation isn’t supported for Intermittent Summaries.

Agent Coaching Analytics on the Dashboard

The Agent AI dashboard now includes an Agent Coaching section with multi-level coaching metrics and rule performance insights to help supervisors monitor adoption and adherence.

Summary Settings

A new Summary Settings sub-section is now available under Agent AI Configurations, centralizing summary preferences for Chat, Voice, and Email with dedicated settings for Landing Summary, Intermittent Summary, End-of-Conversation Summary, and Summary Generation Mechanism.

GenAI Runtime Support for Playbooks

Playbooks now support a GenAI runtime option that uses conversation history to improve trigger detection and step adherence, increasing accuracy for multi-action steps and multilingual conversations. Admins can select Deterministic or GenAI mode when creating a new playbook.

Admins can also configure a context window under Widget Settings to define how many recent messages the system uses for evaluation (default is five). A warning appears when a stage has more than five steps, as additional steps can increase latency.

Custom Model Support for Real-Time Coaching

Real-Time Coaching now supports custom LLM models through the Models Library. Admins can select a custom model for coaching rules and define GenAI prompts to deliver real-time coaching based on business needs. This improves flexibility and coaching accuracy with domain-specific models, aligns outputs with internal standards, and supports security and compliance requirements while reducing dependence on third-party APIs.

Original source

Console

Agent Screen Recording for Voice Calls

Agent screen recording for voice calls is now available, with an app-level toggle and a Windows desktop capture application that records full desktop activity and streams it securely. Recording starts on call acceptance, continues through hold, transfer, and ACW, and stops after ACW submission. Recordings are stored securely, available on the Interactions page for download with role-based access, and adhere to existing retention policies. Validation, error handling, and compliance controls ensure reliable and secure operation.

Warm Transfer for Live Chat

The new Consult option allows agents to initiate an internal agent-to-agent chat, share context, and collaborate without displaying messages to the customer. Agents can switch between customer and consult chats, merge into a conference chat, and complete a warm transfer with full context.

Real-Time Status and ACW Timers

A real-time timer in the Agent Status dropdown shows elapsed time for the current status session and resets on every status change, without affecting cumulative reporting. Each conversation in ACW shows its own timer, starting at ACW entry, running independently, and stopping on closure.

Anchor Tag in User Info

User Info now supports a predefined anchor tag syntax that renders clickable links in the Console. Clicking a link opens the URL in a new tab or window.

Agent-Controlled Skill and Queue Subscription

Agents can now opt in or out of assigned skills and queues from the Profile section when permitted, with deselected items remaining visible but excluded from routing. A Revert to Default option restores original assignments. Supervisors retain full visibility of assignments and agent-level changes, with the ability to override.

Configuration

Queue and Agent Level Blended Agent Configuration

Blended agent configuration is now supported at the queue and agent levels, in addition to the app-level setting. Agent-level override takes the highest priority, followed by queue-level, then app-level. Routing and system busy logic use the effective blending state, ensuring optimized workload distribution while retaining existing channel selection behavior.

Dashboard: Queue-Based Access Control for Supervisors

A new Queues option under Dashboard Content permissions now restricts dashboard visibility to a supervisor’s assigned queues and the agents serving those queues. This applies across dashboards, wallboards, and interactions, including filters, summaries, and searches.

Team Leader Permission Modification

A new Edit User Information child permission under User Management controls who can modify user profiles. Users without this permission can view users and perform operational tasks-routing, queues, and monitoring-but can’t edit agent or user details, with profile pages rendered in read-only mode.

Dedicated Waiting Experience for Email Channel

A separate Email Waiting Experience is now available under Flows, supporting Initial Email, Agent Transfer messages, and configurable periodic messages with hours/minutes and loop/once behavior.

Analytics

Update Agent Status Label:

Deleted →
Inactive

The Deleted label is replaced with Inactive where Agent Name appears across reports, dashboards (Queues, Agents, Interactions), wallboards, filters, and configuration screens now return Inactive instead of Deleted, ensuring consistency across UI and data outputs.

Campaigns

Contacts: List View

A new list view is now available for contacts from CSV, Passive API, and Subset-based sources. The view supports pagination, column selection, exact-match search, priority filtering, sorting, and CSV export. Key contact fields are displayed in read-only format.

API

Check Agent Availability API: Channel-Level Availability and Capacity Details

The API now returns channel-level availability for Voice, Chat, Email, and Messaging, including channel enablement and real availability based on capacity. Each channel includes total, used, and available capacity, along with a digital experience indicator. Availability reflects true readiness only when status, channel enablement, and capacity conditions are all met.

Original source

This update includes new features and feature enhancements summarized below.

Projects

Create self-contained workspaces where you upload files, define custom instructions, and carry context across conversations, so you can ask queries without re-uploading documents or repeating background information every time. Each project maintains its own knowledge base, conversation history, and compose bar scoped to the project’s content.

• Create a project with a name, optional instructions, and file uploads (PDF, DOCX, PPT, TXT up to 25 files, 15 MB each). Instructions guide how responses are generated whenever you query within the project.
• The project detail view displays files, instructions, and recent threads in the left panel, with a project-scoped compose bar, auto-generated smart suggestions, and custom sample queries in the right panel.
• Hover over any file to Ask against it individually, or select multiple files for bulk actions (delete, move, query).
• Project threads appear in History with a project icon and reopen with full project context preserved.
• Share projects with specific users as read-only. Shared users can view files, suggestions, and sample queries, but cannot modify project settings.

Learn more→

Document Generation

Generate documents directly in the chat interface without downloading. Supported formats include PPT/PPTX, DOC/DOCX, XLS/XLSX, and PDF. Mermaid diagrams and HTML code render visually with a toggle to switch between the rendered view and raw source.

Left Panel Navigation Updates

The left navigation panel now includes an Account Switcher at the top, allowing users to switch between accounts they have access to. The switcher is hidden when only one account is available.

The History section displays the four most recent threads inline. A See All option opens a pop-up with all threads grouped by date (Today, Yesterday, and specific dates)

Original source

This update includes enhancements and bug fixes. The key enhancements included in this release are summarized below.

Configuration

Queue Settings: Extended ACW Duration

Queue settings now support configuring After Call Work (ACW) duration up to 90 minutes. Inputs are validated to ensure values fall within 10 seconds to 90 minutes, with errors displayed for invalid entries. This extends configuration flexibility without impacting existing queue behavior or live interactions.

Analytics

Interaction Details Report: Disposition Notes

A new Disposition Notes column has been added to the Interaction Details Report CSV output. It captures agent-modified notes with full support for multi-line content, special characters, and Unicode, without impacting the existing report structure.

Original source

This update includes new features and enhancements summarized below.

Multi-Agent Orchestration

Custom Tags Support

Agent Platform introduces support for custom tags to attach user, session, and message-level metadata to interactions. These tags are included in the Execute API response payload, enabling client applications to consume them for downstream processing and analytics in AI for Service.

Learn more →

Expanded A2A Protocol Support

The Agent Platform now supports JSON-RPC transport binding for A2A v1.0, in addition to the existing HTTP-JSON support. Full support for A2A v0.3 is also introduced, covering both HTTP-JSON and JSON-RPC transport bindings, ensuring broader compatibility and flexibility across integrations.

Learn more →

List Sessions API

A new List Sessions API is now available to retrieve sessions within a specified environment. The API supports filtering by time range, session reference, and user reference to help narrow down results. If no environment is provided, sessions are retrieved from the draft environment by default. This helps streamline session discovery and simplifies troubleshooting and analysis.

Learn more →

Expanded Real-time V2V Model Support for Adaptive Network

The Agent Platform now supports additional real-time models for Voice-to-Voice (V2V) interactions within the Adaptive Network Pattern framework, extending beyond existing OpenAI support to include:

• Azure OpenAI
• Grok
• Ultravox

Note: Gemini models aren’t supported in Adaptive Network.

Learn more →

Playground Enhancements

The enhanced Playground significantly improves the testing and validation experience for developers:

Inline Artifacts Support:

Playground now supports inline rendering of artifacts, allowing developers to preview outputs without full deployment. When enabled, artifacts appear within the conversation flow, improving visibility and simplifying the validation of structured data. AI for Service-supported templates are rendered using native UI components, while non-supported artifacts are displayed as raw JSON.

Learn More

Configurable Session Metadata:

Playground now supports session-level customization, allowing users to configure metadata and key runtime settings directly within the testing interface. Developers can provide metadata to simulate a real runtime context and configure settings such as artifacts display, streaming, thought streaming, and document upload—eliminating context switching and enabling more efficient testing. This feature is currently in preview and can be enabled upon request.

Configurable Transcription for Realtime Sessions

The Platform now supports configuring transcription language and prompts for real-time voice-to-voice sessions, improving the accuracy and efficiency of speech-to-text processing. These settings, applied at the app level, allow developers to specify the input language and provide domain-specific context, enabling more accurate transcriptions and better recognition of specialized vocabulary. By default, the ASR model operates in autodetect language mode.

Learn more →

AI Engineering Tools

Expanded Model Support

The Platform now supports additional AI models, including:

• OpenAI : gpt-5.4, gpt-5.4-nano, gpt-5.4-mini, gpt-realtime-1.5.
• Azure OpenAI : GPT-5.3-Chat, GPT-5.4, GPT-5.4-Nano, GPT-5.4-Mini, GPT-Realtime-1.5.

Learn more →

AI Safety, Security, and Governance

Data Anonymization: Module-Level Control for Anonymization and Deanonymization

The Platform introduces a new unified guardrail framework that consolidates PII Guardrails and Anonymization / Deanonymization into a single configuration interface. Entities are defined once per entity, using either regex-based PII detection or ML-based anonymization, with both layers executing in a unified processing sequence across all platform stages.

Access controls are now configurable at the module level, specifying whether users, code tools, workflow tools, MCP tools, events, pre-processors, and proxy agents receive original or redacted values at each processing stage.

Learn more →

Native mTLS Support for OAuth 2.0 Client Credential Auth Profiles

OAuth 2.0 Client Credential auth profiles now support mutual TLS (mTLS) natively, enabling secure connections to systems that require mTLS without external tools or custom workarounds. The platform can present a client certificate for both token requests and API calls, ensuring compatibility with enterprise systems that enforce mTLS.

Learn more →

Original source

This update includes new features and feature enhancements summarized below.

Employee Organization Chart

The Employee Directory now includes an Organization Chart tab that displays the organizational hierarchy visually. Expand employee cards to view direct and indirect reports, apply filters, and search for specific employees. The chart automatically focuses and expands to the matched user. Open the full profile view from any card. Zoom, fit preview, collapse, and full-screen controls are also available.

Learn more→

Feedback Management and Email Notifications

Introducing a dedicated Feedback page under Usage in the Admin Console. Administrators can now view, filter, and drill into user feedback across all agents and workspaces, with access scoped by role. Account-level admins see all feedback across the account, Workspace Owners see feedback within their workspace, and agent-level access scopes feedback to that specific agent.

End users can now choose from an expanded set of feedback categories when submitting upvote or downvote feedback, including Forgot previous context, Outdated, Wrong formatting, and Other, in addition to the existing options.

Learn more→

UI Localization and Language Support

AI for Work now supports multi-language localization. On first login, users can select their preferred language from the language selection module. All static and dynamic UI text updates immediately across the application. Users can change their language preference at any time through the profile icon.

Arabic includes full RTL layout support. All UI elements, including the history panel, response area, menus, and icons, realign to RTL automatically.

Learn more→

Drag-and-Drop File Upload

Users can now drag and drop files directly into the AI for Work chat interface, in addition to the existing attachment icon workflow. The drop zone highlights on hover, and the upload initiates automatically once a file is dropped.

Learn more→

Scoped Source Selection for Search Agents

When configuring a Search Agent or adding a Search App to Enterprise Knowledge, users can now select or deselect specific sources, including websites, file uploads, and connectors, from the linked Search AI app after authentication. Previously, all available sources connected by default. Previously selected sources are retained when editing credentials or sources.

Learn more→

Search Filters

Users can now filter search results in the search results panel by Timeline, Created By, and File Type. The Timeline filter supports predefined ranges and a custom date picker. Created By supports multi-user selection with a searchable pop-up, and File Type options populate dynamically based on the current result set.

Original source

This update includes new features and enhancements summarized below.

Multi-Agent Orchestration

Voice-to-Voice Support for Adaptive Network

Voice-to-voice models are now supported in the Adaptive Network, enabling seamless processing of spoken input and generation of spoken responses. This enhances conversational experiences by enabling more natural, real-time voice interactions.

Complete App Export and Import

Agent Platform now supports full application export, packaging all components, including workflow tools, into a single file for seamless migration across environments.

The import process includes upfront validation before execution and automatic rollback on failure, ensuring imports either complete fully or not at all. This eliminates the risk of partial or inconsistent application states after a failed import.

Learn more →

Pre-Processor Execution Control

Users can configure execution control for pre-processors, choosing whether they run once per session or on every agent invocation. This reduces latency and avoids redundant processing. Existing configurations default to Always Run, ensuring backward compatibility.

Learn more →

Response Processors for Output Transformation

Agent Platform introduces the Response Processor, a new capability that gives full control over how responses are shaped and delivered across channels. This feature enables channel-based, structured responses via templates, allowing you to define the exact response format for each channel. Admins can modify the existing artifacts key to reshape the output on the fly, or update it entirely with a customized structured response tailored to the target channel. Developers can further apply custom formatting, enrichment, and business logic via code, with full access to the response context, including inputs, outputs, and artifacts, all without changing the underlying logic.

Learn more →

AI Safety, Security, and Governance

Expanded Model Support

The Platform now supports additional AI models, including:

• OpenAI: gpt-5.3-chat-latest
• Anthropic: claude-sonnet-4-6
• Grok Realtime (Available via custom integration)

Learn more →

Original source

This update includes enhancements and bug fixes. The key enhancements included in this release are summarized below.

UI Enhancements

Pass Metadata Support in Search AI Configuration

In the newly available Pass Metadata configuration section, you can select Landing Summary, Agent-Customer Chat History, and User Context to pass as metadata. This improves response relevance, personalization, and accuracy without requiring manual query rephrasing.

Spanish Support for the Agent AI UXO App Language

Support for Spanish as an App Language is now available via profile menu > app language.

Integration Enhancements

Custom Field Support for Salesforce Email Channel

Agent AI integration with the Salesforce Email Channel now supports custom field objects. Salesforce admins can configure custom data in the Agent AI Configuration page by mapping payload keys to Case field API names using a .json file. The system resolves these fields at runtime and sends case data to the Agent AI widget.

Agent AI integration with Five9 Voice

Agent AI integration with Five9 Voice supports Automation, Agent Coaching, Agent Playbook, Sentiment Analysis, Transcription, end-of-call summary, and Custom Data. It also supports Secure Custom Data transfer, such as agent name and ID, to the Agent AI widget.

SDK Enhancement

Headless SDK for Cross-Platform Integration

The new headless SDK for browser and Node.js environments, with support for standard loading methods, enables message handling, internal transfers, summary generation, and conversation closure through a unified API. It removes UI dependencies and simplifies integrations.

Original source

Configuration

Separate Wait Time Controls for Post-Transfer Conversations

Conversations now use separate queue wait time controls before and after agent transfer. On the first entry, the initial wait time applies. After a transfer, a dedicated post-transfer timer activates and resets with each subsequent transfer.

Parent and Child Grouping of Permissions

Two new permissions, in the Monitor Tab, are Profile and Mark Offline. These improve role-based access control and will be enabled by default for all roles except Agent and Custom.

Channel-Based Last Agent (Affinity) Routing

Last Agent Routing is now available per channel (Voice, Chat, and Email), routing interactions to the last handling agent when available. An optional wait duration holds the interaction before falling back to standard routing; if no wait duration is set, fallback occurs immediately.

Analytics

Correct Call Status for Agent Interactions with CSAT

The system now records accurate call-termination status during agent-customer interactions, including CSAT. Calls are marked as User Hangup or Agent Hangup, preventing incorrect classification as Bot Hangup.

Show Duration Markers in Recordings

A new account-level setting lets you display either timestamps or elapsed call duration for transcript segments. When duration is selected, each segment shows time relative to the call start instead of the date and time. Timestamps remain the default, and the setting applies to all recordings on reload.

Usability Enhancement to View and Navigate Recordings

The Insights to Logs panel now includes a compact recording overview widget, an expandable full-screen view, and a zoom control, which together improve recording visibility and navigation.

Campaigns

Add Payload Input for API-Integrated Contact List

POST-based API-integrated contact lists now support request payload input. The payload field accepts .json file, raw text, or URL-encoded data, with existing validation rules applied. This enables flexible data exchange with external systems without affecting current API configurations.

Agentless Dialers: Disposition Support

Agentless voice campaigns now support predefined disposition codes for tracking and analytics. The system attaches the selected disposition to the session as a tag and displays it in campaign outcomes and interactions.

New Dashboard for Proactive Web Campaigns

A dedicated dashboard for proactive campaigns is now available, with filters for campaign executions and date ranges. Key metrics include impressions, clicks, visitors (unique, anonymous, and logged-in), conversations, and dismissal rate. Visualizations cover time trends, reasons for suppression, conversation split (agent vs. bot), and disposition distribution.

API

Time-Based Queue Reprioritization

A new public API lets you dynamically update queue priorities, enabling time-based automation via external schedulers. Updates apply immediately, override UI-configured priorities, and influence routing without reassigning active conversations.

Original source

This update includes enhancements and bug fixes. The key enhancements included in this release are summarized below.

Integration Enhancements

Open Knowledge Base Articles Within the Salesforce Workspace

The Agent AI widget in Salesforce now allows agents to access knowledge base (KB) articles directly from the Search tab within a workspace subtab of the current browser session. This keeps agents in the active session, preventing reauthentication prompts or interruptions to the customer interaction.

Automatic Conversation Summary After Einstein Bot Handoff in Salesforce

The Agent AI widget in Salesforce now automatically generates a conversation summary when a customer interaction is handed off from Einstein Bot to a live agent. The summary is displayed in the Assist tab, giving agents quick context on the customer’s issue and actions already taken — without reviewing the full chat transcript. The summary is also stored in interaction metadata for downstream use.

Automatic Summary Updates for Email-to-Case in Salesforce

The Agent AI widget for the Email-to-Case channel in Salesforce now keeps the Assist tab summary up to date with the latest email conversation. When the widget loads, it displays a summary of the existing email thread. As new emails arrive or agents generate or submit a summary, the Assist tab updates to reflect the latest context.

UI Enhancements

Expand and Collapse Controls for Customer Messages and Agentic NBA Suggestions

In the Agent AI widget’s Assist tab, customer messages and Agentic Next Best Action (NBA) suggestions now include expand and collapse buttons that appear on hover. By default, the messages and the previous suggestions (sent or unsent) are collapsed, while the latest suggestion appears expanded. This prevents messages from automatically collapsing while agents copy suggestion content.

Original source