Changelog
MaiAgent platform updates and new feature release notes
v2026.09.07
✨ New feature release
AI Gateway and model management updates: This release delivers 1 set of improvements now live, covering “AI Gateway and model management updates”.
Agent Marketplace and AI assistant updates: This release delivers 1 set of improvements now live, covering “Agent Marketplace and AI assistant updates”.
AgentOps and quality monitoring updates: This release delivers 1 set of improvements now live, covering “AgentOps and quality monitoring updates”.
MaiGPT, Canvas, and document preview updates: This release delivers 1 set of improvements now live, covering “MaiGPT, Canvas, and document preview updates”.
Web Chat and Ask AI updates: This release delivers 1 set of improvements now live, covering “Web Chat and Ask AI updates”.
Platform performance and stability updates: This release delivers 1 set of improvements now live, covering “Platform performance and stability updates”.
Meeting notes updates: This release delivers 1 set of improvements now live, covering “Meeting notes updates”.
Usage, Credit, and billing updates: This release delivers 1 set of improvements now live, covering “Usage, Credit, and billing updates”.
Knowledge base, file, and search updates: This release delivers 1 set of improvements now live, covering “Knowledge base, file, and search updates”.
Organization, role, and permission updates: This release delivers 1 set of improvements now live, covering “Organization, role, and permission updates”.
Connectors, tools, and third-party integration updates: This release delivers 1 set of improvements now live, covering “Connectors, tools, and third-party integration updates”.
🔐 Security and governance enhancements
Brand and shared interface updates: This release delivers 1 set of improvements now live, covering “Brand and shared interface updates”.
🛠️ Bug fixes
Database and contact updates: This release delivers 1 set of improvements now live, covering “Database and contact updates”.
v2026.09.04
✨ New Features
Brand-new Work workspace: A new Work workspace brings assistant switching, cross-assistant conversation history, and frequently used entry points together in a single three-zone screen. The system remembers your last-used workspace and layout, and when you land on a page you don't have permission for, it automatically takes you to a feature you can use.
Workspace permission preview for roles: The MaiAgent role editor page adds workspace permission cards and a live preview, so you can confirm which functional zones a role can access before saving — reducing cases where a misconfigured permission is only discovered afterward.
File library integration with AI workflows: The file library can now show AI analysis status, mount files into knowledge bases, and bring files into the MaiGPT workspace. Once an attachment finishes parsing, it is placed directly into the working context, and for long documents the most relevant content is provided within the available space.
Multilingual content for built-in agents: The 98 built-in agents in the Agent marketplace now display their names, descriptions, and interaction content in English or Japanese according to the interface language, so switching languages no longer leaves content mixed with other languages.
AI Slides automatic image generation: AI Slides can now generate accompanying images as part of the creation flow and place the generated images directly into the slides, with no need to download them first and re-upload.
Question cards support skipping: When an AI assistant asks a multiple-choice question, non-essential questions can now offer a "Skip" action, so users no longer have to pick an ill-fitting answer just to keep the conversation going.
Deep research file downloads: When a deep research run completes, you can now obtain a downloadable report file directly, making it easy to save, share, or continue editing the research results.
🚀 Core Performance Improvements
More stable test dataset evaluation: Large test dataset runs are now split across multiple workers that claim tasks on demand. If an individual worker is interrupted, its work is automatically reclaimed and retried, and completed cases persist their actual usage — so long-running evaluations no longer stall because a single worker was interrupted.
Faster visual parsing of large documents: Long documents containing images or complex layouts are now parsed in concurrent batches. When the service is busy, processing automatically slows down and continues, shortening wait times and reducing whole-document failures.
Faster organization usage page loading: Organization usage and quota summaries are now aggregated in a leaner way, so administrators get the complete statistics faster when opening the usage page.
Faster member roster loading: When a large organization loads its full member roster, the system now prepares exactly the data the screen needs, reducing the wait caused by processing large numbers of members one by one.
Faster channel knowledge summaries: Channel knowledge summaries and select-all data now load faster with large numbers of documents, shortening the wait when configuring knowledge scope.
Faster SCIM group synchronization: When an enterprise identity directory syncs group members, only the members actually added or removed in that update are processed — making large group updates faster and lowering the load during synchronization.
Faster LINE message ingestion: When a LINE channel receives consecutive messages, it now reuses still-valid user profile data, reducing the wait caused by repeatedly querying LINE so messages enter processing faster.
😊 User Experience and Interface Improvements
Login page layout alignment: The loading indicator on the enterprise login page is now centered, and the fields in the account-detection and password-entry steps are laid out consistently, so the screen stays stable when moving between login steps.
Workflow reply mode retirement: AI assistants no longer offer creating new Workflow reply modes, and duplicating an older assistant now switches it to standard Q&A. Existing assistants that have gone unused for an extended period are automatically converted back to standard Q&A, while recently active ones keep their current behavior.
🛠️ Bug Fixes
Activity log system event display fix: Fixed an issue where automatic system events misaligned the columns in the activity log; even when an event has no operator, the date, action, and content now line up in the correct columns.
AI monitoring log loading fix: Fixed an issue where the AI monitoring log page occasionally failed to load; the page now fetches data only after the table is ready, so it no longer shows up blank due to initialization order.
MCP App teardown stability: Fixed an issue where closing an MCP App card or switching conversations occasionally produced errors; finished connections are now cleaned up properly and no longer interfere with subsequent conversations.
MCP App tool timeout notice: Fixed an issue where MCP App tools could hang without responding; if a third-party service does not reply within 30 seconds, the system now stops waiting and shows the execution as failed, so users can retry immediately instead of waiting several minutes.
Web Chat sub-path deployment fix: Fixed an issue where Web Chat deployed under a reverse proxy sub-path failed to open properly; the system now determines the path from the actual URL, so refreshing and navigating between pages both work correctly.
Skill export and re-import fix: Fixed issues where an exported Skill could fail to import back or lose settings; re-importing now preserves the usage timing and bound tools, and names are automatically adjusted to an importable format.
Ask AI organization scope fix: Fixed an issue where Ask AI could keep using the previous organization's data scope after switching organizations; the tools now always run against the organization the current conversation belongs to, avoiding irrelevant cross-organization results.
Usage statistics sorting fix: Fixed an issue where sorting usage statistics by credits only ordered the current page; the full result set is now sorted before pagination, so the order stays correct when browsing across pages.
Knowledge base search status notice: Improved the message shown when a knowledge base search returns no results; if relevant documents are still being processed, the AI assistant now points out the files that are not yet ready — including workspace-mounted content — so users don't mistakenly conclude the knowledge base is empty.
Error message for models without tool support: Improved the error message shown when a model does not support tool operations; the screen now clearly explains that the current model cannot complete tool tasks and guides users toward a supported model, instead of only showing "Please try again later."
Automatic fallback for image generation: Strengthened image generation stability; when the primary service has a temporary outage, the system automatically switches to a fallback service, and if a credit deduction does not complete, it is clearly reported instead of interrupting with an unexplained error.
Schedule execution limit notice: Fixed an issue where schedules that reached their execution-count limit paused over and over; the system now clearly states that the available executions have been used up, making it easy for users to adjust the schedule settings.
Anonymous Web Chat output file downloads: Fixed an issue where anonymous Web Chat could not open files and images produced by the code interpreter; visitors can now view or download the results directly in the original conversation without being redirected to log in.
Monitoring timeline step fix: Fixed an issue where the AgentOps trace timeline was missing model-processing steps; administrators can once again see every model execution segment, making it easier to pinpoint reply quality and latency issues.
Long-question knowledge retrieval fix: Fixed cases where very long questions caused knowledge retrieval to fail; the system now keeps the usable question content within the service limits, so queries with large amounts of pasted background material can still complete.
Meeting participant access fix: Fixed an issue where participant permissions were incomplete for meeting records created from calendar events; the system now identifies internal and external participants from the invite list, so everyone who should have access can view the meeting content normally.
v2026.09.02
✨ New Features
Real-time meeting backup recording prevents audio loss: Real-time meeting records now include local backup recording: audio is saved locally in the browser before uploading, so even if the network drops or the page is accidentally closed mid-meeting, the recorded content is automatically uploaded once you reconnect. The meeting detail page lets you switch between the original and backup audio files and re-transcribe using the backup audio, and the list marks meetings whose backup is still syncing or incomplete.
Self-service management of "Connected Services": The conversation pages of Web Chat and built-in AI assistants now let you manage third-party service authorizations yourself: click the "Connected Services" entry at the top of the page to view the authorization status and validity of services such as calendars and cloud drives, and authorize, re-authorize, or disconnect directly (with a confirmation before disconnecting). Authorization requests received during conversations are now presented as clear authorization cards instead of a string of raw English URLs, and the mobile layout works the same way.
Login page automatically detects the SSO login path: The login page now determines the login method automatically based on the email you enter: enterprise SSO users are redirected straight to their corporate single sign-on after entering their company email, with no special URL to remember, while regular users enter their password as usual. When SSO validation fails, a clear error is shown with options to retry or switch to password login, instead of getting stuck in a redirect loop.
Webhook channel added to human handoff notifications: Human handoff notification settings now support a Webhook channel. You can enter your own system's HTTPS receiving URL (with an optional signing secret so the receiver can verify the source), and when a conversation is assigned for human handoff, the platform pushes an event notification to that URL in real time — making it easy to integrate the handoff flow with external systems such as ticketing or on-call alerting. Failed deliveries are retried automatically, and existing Email and notification center notifications are unaffected.
Hierarchical channel-type filter added to the contacts page: The contacts page adds a two-level filter: first select the channel type (web, LINE, Telegram, Facebook, etc.), then select a conversation platform under that type, and the list narrows in real time; the conversation platform dropdown also shows channel labels, making it faster to find contacts across platforms.
Batch credit quota deactivation can also release restrictions: When batch-deactivating credit quotas by MaiAgent role, you can now choose to also release members' quota restrictions and switch them to the organization wallet, preventing members from being locked in a "restricted with no quota" state; the confirmation dialog also shows the accurate counts of members to be deactivated and to be released, based on actual statistics.
🚀 Core Performance Improvements
Built-in Agent management page no longer lags when switching: The built-in Agent management page no longer re-queries each Agent's access targets when switching categories or searching. Organizations with many activated Agents will notice much faster tab switching, and the page no longer flashes a full-page loading screen.
Faster opening of conversations with massive tool output: Opening old conversations containing large amounts of tool execution results no longer freezes, and memory usage while typing drops significantly (measured from 255MB down to 158MB); tool step details now load only when expanded, and extremely long results show the beginning with a truncation notice.
Faster loading of knowledge base full-text view: The knowledge base's full-text document view now assembles faster when opening large documents (hundreds of chunks); chunk mapping uses a more efficient algorithm, with the displayed content completely unchanged.
Improved platform stability during system updates: During system update deployments, the momentary server load caused by mass Web Chat connection reconnects has been greatly flattened, reducing platform-wide slowdowns during update windows; users waiting for a reply are unaffected and still receive messages in real time.
😊 User Experience and Interface Improvements
Comprehensive improvements to conversation scrolling and reading: The conversation scrolling experience across Web Chat, the admin console chat, and MaiGPT has been comprehensively improved:
While the AI is generating an answer, you can freely scroll up to review earlier messages without the view being forcibly pulled back to the latest position;
The paragraph you are reading no longer drifts upward as the reply generates, and after loading more history the view stays at your original reading position;
All conversation interfaces now provide a unified "back to latest" button — click it to return to the latest content and resume auto-follow;
The opening greeting message is no longer pushed off-screen.
Redesigned conversation platform settings page navigation: The conversation platform settings page has been redesigned to match the AI assistant settings page layout: settings categories move to a grouped left sidebar, while Basic, Quota, and Auto-reply become top tabs; category switches are recorded in the URL, so refreshing or sharing the link returns to the same category.
MaiAgent role permission checkboxes take effect instantly: When editing a MaiAgent role's resource permissions, checking any permission now takes effect immediately with a result notification — no more pressing "Update Role" tab by tab, and no more losing some changes after editing across tabs.
Cache items split out in credit consumption breakdown: The conversation credit consumption summary now lists "LLM cache read" and "LLM cache write" as separate items instead of lumping them into "Other", so cache costs are clearly visible when using models with automatic caching such as Gemini; category names in narrow panels now wrap to a new line instead of being truncated.
Slide and document templates can be deleted: When using slide or document templates in a conversation, templates uploaded by your organization can now be deleted directly: template cards gain a delete button that asks for confirmation before deleting, so mistakenly uploaded templates no longer require an engineer to remove.
Agent Marketplace shows required connectors and setup hints: In the Agent Marketplace and the built-in AI assistant management page, cards and preview windows now display the third-party service connector icons the AI assistant requires; assistants that need an administrator to complete connection setup first are labeled "Administrator setup required", preventing them from being mistaken as ready to use out of the box.
Built-in Agent usage renamed to "activation count": The built-in AI assistant usage wording on the organization management page changes from "seats" to "activation count", clearly conveying that this is the maximum number of AI assistants that can be active at once, no longer confused with per-user licensing; all five locales — Traditional Chinese, Simplified Chinese, English, Japanese, and Korean — are updated in sync.
Voice conversation replies now render with full formatting: In Web Chat voice conversation mode, the AI assistant's text replies now render with full formatting — lists, headings, and quotes are laid out properly and left-aligned, with no more raw symbols or misaligned centering.
Unified sidebar brand logo sizing: The brand logo in the admin console sidebar now uses a fixed size and centered display, making logos from different brands appear more consistent, with no distortion in either the expanded or collapsed state.
🔐 Security and Governance Enhancements
Embed widget security hardening and integrity verification: Customers embedding Web Chat on their own websites can now enable stricter website Content Security Policies (CSP) with the chat button and window appearance completely unaffected; version-pinned embed files with integrity verification information are also provided to satisfy security scanning and audit requirements. The chat button's hover zoom effect also takes effect properly as of this release.
Tightened access to the built-in Agent management entry: The built-in Agent management entry is now controlled by explicit authorization via the "Built-in Agent management permission" under organization permissions: members without this authorization no longer see the management entry or the marketplace's "Enter management" link, while marketplace browsing is unaffected — making permission management more precise.
Permission gating on applying Agent Marketplace templates: The "Use this Agent" button on Agent Marketplace detail pages now respects permissions: members not granted marketplace usage permission see the button grayed out with a tooltip, instead of discovering they cannot apply the template only after clicking. Identity zone derivation is also fixed so members with only conversation permission are no longer misclassified as builders.
🛠️ Bug Fixes
Vision parsing engine upgraded and blank-page issue fixed: When parsing documents with the vision parsing engine (Vision Parser), the issue where some pages could be silently parsed as blank — causing gaps in knowledge base content — has been fixed; the parsing engine's underlying model is also upgraded to Claude Sonnet 5, with parsing quality tested across Chinese, English, Japanese, and Thai to be on par with the previous version. Additionally, due to upstream service quota limits, the MaiAgent Parser (Online) parser has been disabled platform-wide.
Fixed Web Chat crashing on open on some Android devices: Opening Web Chat in certain Android in-app browsers (environments with local storage disabled) no longer crashes the entire page, and chat features work normally.
Fixed login failure on on-premises deployments: The inability to log in to Web Chat and the admin console in private (on-premises) deployment environments has been fixed, while also ensuring that frontend error information from on-premises environments is never sent back to the cloud — data stays entirely within the customer's environment.
Fixed scrambled ordering in multi-file uploads: When uploading multiple images or files at once in Web Chat or MaiGPT, attachments now stay in the order you added them: the AI assistant reads images in the same order as the on-screen thumbnails, so asking about "the first image" returns the right one; files still upload in parallel, so speed is unaffected.
Tool execution narration follows the interface language: After switching the interface to English, Japanese, or another language, system default tool narrations in Web Chat and admin console conversations — such as "Searching knowledge base" and "Searched the web" — now display in the interface language, and narrations in existing conversations update in sync after switching; administrator-customized narration text is unaffected.
Fixed blank screen after leaving the conversation log detail page: Fixed the issue where, after viewing a conversation log's details in AI assistant monitoring, switching to another page left the content area completely blank and required a refresh to recover.
Fixed read-only conversation links requiring login: Anyone with a conversation view link can once again view conversations that allow public read-only access without logging in; the login prompt appears only when the conversation genuinely requires permission, and ordinary connection issues show a loading failure instead of misleadingly redirecting to login.
Fixed out-of-sync list previews after editing a message: After editing a message in the conversation platform, conversation list previews in other tabs and on other support agents' screens now show the edited content in real time without a manual refresh, and the conversation order does not change as a result.
Fixed inaccurate AI assistant quota display: The AI assistant card in the organization usage overview now shows your plan's true limit and actual usage count: organizations with purchased add-on quotas see the correct denominator, legacy-mode assistants retained after a downgrade are clearly marked as not counting against the quota, and the card tells you directly how many more you can add.
Fixed built-in Agent usage numbers changing with filters: The usage card on the built-in Agent management page now consistently shows the organization's actual total activation count; the number no longer changes when searching or filtering the list, and matches the quota information in the bottom batch-action bar.
Fixed MaiGPT page stuck half-initialized: When opening the MaiGPT conversation page, even if the model list temporarily fails to load or times out, the page now finishes initializing and conversations work normally — no more getting stuck in a half-loaded state demanding a refresh, and wait times are greatly reduced.
Fixed momentary display errors when loading MaiGPT conversations: When entering the MaiGPT conversation page after logging in, the message list no longer briefly fails to display while assistant data is still loading; failed login form validation keeps the inline red field hints without generating extraneous background errors.
Fixed workspace entry appearing incorrectly with errors: Fixed the "File Workspace" option in MaiGPT conversations being shown to organizations that have not yet enabled this capability: organizations without it no longer see an entry that only produced permission errors when clicked, while organizations with it enabled are unaffected.
Fixed AI Gateway overview not scrolling: Fixed the AI Gateway overview page being unable to scroll down in shorter windows: content exceeding the screen can now scroll normally, and bottom sections such as "Usage per API Key" are fully viewable.
Fixed misaligned rows in the activity log: When the activity log in organization settings displays system or organization events, columns now align correctly — the user column is left blank, and the view button and table zebra striping are no longer scrambled.
Fixed submission failure for overly long personal preferences: The character limit for personal preferences is adjusted to 1,000 characters, matching the server's actual limit; long input is no longer rejected only at submission time.
Fixed clipped buttons in the API key list: When horizontally scrolling the table on the API key management page in organization settings, the action buttons on the right now display in full instead of being clipped and hard to click.
File library menu position and space name locale fixes: The "File Library" item in the left menu moves below "Agent Marketplace", better matching usage flow; file library space names such as "My Files" and "Organization Shared Files" now follow the interface language instead of always displaying in English.
Fixed guest draft attachment removal not working: After uploading an attachment as a guest in Web Chat, clicking × to remove it before sending now actually takes effect, and the AI assistant no longer references removed files; if removal fails, a notification appears. Also fixed removing an attachment possibly deleting a different one when multiple file types were uploaded at once.
Tool card display setting no longer lingers: Fixed tool execution cards being permanently hidden and lingering across conversations after switching from a brand animation conversation; tool cards now display correctly according to settings every time you enter a conversation.
Fixed oversized brand logos: The brand logo sizes at the bottom of the Web Chat conversation page and on the login, sign-up, and password reset pages have been corrected to reasonable dimensions, and the splash screen logo is no longer blurry on high-resolution screens.
Fixed connection cleanup interruption when closing conversations: The internal cleanup flow when logging out of or closing Web Chat is more stable, no longer failing midway due to repeated connection disconnects, and avoiding leftover conversation or authentication state.
Last updated
Was this helpful?