V3.226.0603 (2026-06-03)

FlowAgent - Classifier

• The FlowAgent Classifier now supports Passthrough mode. The Classifier currently provides two modes: Passthrough and Extraction. In Passthrough mode, the Classifier passes the user's original input text directly to the activated branches (which may be one or multiple). In Extraction mode, the Classifier splits and extracts the user's input according to predefined classification rules and passes the extracted results to the activated branches, so that different branches receive different content.
• Example scenario:

                      
                      User Input: Where should we travel tomorrow? Was lunch tasty today?
Branch Rules: Branch A: Travel-related questions | Branch B: Food-related questions
Classification Mechanism: Multi-branch triggering; one or multiple branches can be triggered based on semantic match.
After Running: Both branches are triggered: Branch A + Branch B
Extraction Mode:
- Message received by Branch A = Where should we travel tomorrow?
- Message received by Branch B = Was lunch tasty today?
Passthrough Mode:
- Message received by Branch A = Where should we travel tomorrow? Was lunch tasty today?
- Message received by Branch B = Where should we travel tomorrow? Was lunch tasty today?

                    

User Attributes - List Type

• List-type fields in User Attributes now support adding and removing each element individually in Debug conversations, improving flexibility in debugging scenarios.

V3.222.0529 (2026-05-29)

ChatGather

The FlowAgent ChatGather component now supports global memory capabilities for Agents, including long-term memory, short-term memory, and key events. User attributes are disabled by default to prevent them from interfering with the quality of data collection tasks.

Agent Tasks

• A new TASK feature has been added to both Agent and FlowAgent, enabling the creation and management of tasks for automated execution and tracking.
• Task trigger methods include three types: Active - Scheduled, Active - Event, and Passive - User Triggered, allowing tasks to be executed in different scenarios.
• Active – Scheduled Trigger: Supports time-based task triggering, including five scheduling options: Once, Daily, Weekly, Monthly, and Interval, to accommodate different execution frequencies.
• Active – Event Trigger: Supports event-based task triggering. Tasks can be triggered via events received through Webhook, enabling execution based on various external events.
• Passive – User Trigger: Supports task triggering based on user behavior, including total number of user conversations and time of the user’s last conversation, enabling tasks to be executed according to different user activity conditions.

OAuth Tools

• OAuth tool support has been added. Currently integrated tools include Google Calendar, Google Mail, Google Docs, and Google Sheets.
• Developers can also configure custom Google OAuth authentication settings, including Client ID, Client Secret, and email, to meet authentication requirements when integrating Google OAuth tools.

Database API

• Create Table API: Supports creating database tables, defining table structures and field attributes. After successful creation, a table ID will be returned.
• Query Table Data API: Supports querying table statistics using the table ID, including the number of fields, data volume, table name, and table description.

V3.218.0521 (2026-05-21)

Form Collection

• The FlowAgent Form Collection component window can now be synchronously displayed in the LiveDesk Widget. End users can complete and submit the form to the AI Agent directly from the LiveDesk Widget, improving customer service efficiency and experience.

Knowledge Base

• Added ACL permission control to the Knowledge Base, supporting permission management across dimensions such as users, departments, and roles, to meet enterprise needs for knowledge base security and access control.
• Added metadata support to the Knowledge Base, allowing custom metadata tags to be added to documents. A custom metadata filtering mechanism is supported in Agent, FlowAgent, and Workflow for precise retrieval of knowledge base documents.
• Added a Knowledge Graph feature to the Knowledge Base, supporting the extraction of uploaded knowledge documents into entity triples to build a knowledge graph at the knowledge base level.
• The knowledge retrieval feature has been upgraded to dual-channel retrieval combining vector search and graph search, improving the accuracy and richness of knowledge retrieval.

V3.209.0511 (2026-05-11)

Integration

• WhatsApp by EngageLab and WhatsApp by Meta now support segmented output, and WhatsApp users can send multimedia messages to GPTBots Agent.

V3.197.0424 (2026-04-24)

FlowAgent

• Optimized the Message Pass-through component: once the component is triggered, users will still receive the message normally even if the Agent's response is terminated.

Agent Configuration - Welcome Guide

• The Welcome Guide now supports multi-language configuration. Once configured, the welcome guide is displayed in the language matching the user's locale.
• If the user's locale has no configured welcome guide, the English welcome guide is displayed by default.

Agent - Users & Key Events

• Added the User List and Key Events features, allowing you to view users who have triggered key events along with their key event details.
• Supports customizing key event types and trigger conditions. When the conditions are met, key events are automatically extracted and generated from conversations.
• Key Events can be enabled in LLM-driven components, allowing the LLM to quickly understand the current user profile and historical key event information from conversations.
• Key Events can be referenced as global variables and used as conditions in if/else branches.

V3.196.0420 (2026-04-20)

Workspace 3.0

Core Capabilities

• Agent Loop Engine: multi-turn autonomous reasoning with circuit breakers, streaming output, and session snapshot & restore.
• Three-tier Memory System: account-level / enterprise-level / session-level memory, multi-level drill-down on the Neo4j knowledge graph, with automatic extraction and injection.
• 5-layer Context Compression: tool truncation → compaction → pruning → LLM summarization → degradation notice.
• Multi-node Collaboration: Gateway smart routing (LLM → BM25 → fallback), supporting "device A chats, device B executes".
• A2A Protocol: structured responses (text / file / data / a2ui), with support for third-party agent integration.

Workspace (for end users)

• Claw Chat: streaming messages, session management, preview of 9 file types, and multimodal input (image/video/document/audio).
• Sub-agents: auto-created + pre-registered, up to 3 levels of nesting, with cross-node dispatch support.
• AI Search: smart search is live; the next release will upgrade to enterprise-grade knowledge search supporting autonomous queries and TB-scale knowledge base lookups.
• Agent Page: A2A agent management and conversation.
• Workflow: workflow orchestration and execution.

APP Settings (APP-only)

• Model configuration (multi-provider API keys + capability tags).
• 21 preset skills + custom + enterprise skills.
• 19 built-in tools + MCP (stdio / SSE / HTTP) + API tools.
• Sub-agents can be pre-created, and the main Agent can also auto-create sub-agents as tasks require.
• 14+ channel integrations: WeChat, Slack, WhatsApp, DingTalk, Teams, Telegram and other IM channels for task execution.
• Account-level memory + knowledge graph visualization.
• Runtime security: contacts, nodes, sandbox security mode, and tool permission levels.

Space Management (for admins)

• Usage statistics (full-dimensional Claw / Search).
• Real-time monitoring and filtering of Claw nodes.
• Unified management of enterprise resources (Skills / Tools / A2A / Agent / Workflow).
• Log auditing (Claw / Search / A2A).
• Safety guardrails (tool permissions / content moderation / keywords).
• Fine-grained member permission authorization, with both feature-level and data-level access control.
• Advanced settings (brand / integration / enterprise memory / theme).

V3.194.0408 (2026-04-08)

Conversation Window Optimization

• Optimized the interaction and display of the time picker in collection forms.
• Collection forms now support the Enter key for submission.
• When the AI outputs long text, the page no longer auto-scrolls to the bottom; instead, a "scroll to bottom" button is shown for users to click.
• On iOS devices, input fields are no longer auto-zoomed when filling in forms; the field is automatically centered within the visible area instead.

A2UI Protocol Support

• The conversation page now renders the A2UI protocol.
• Prompt instructions can ask the LLM to generate A2UI protocol JSON, which the frontend will render as cards, options, input fields, carousels, video, audio and other A2UI components. Users can interact with these components, and the interaction results are resubmitted to the Agent as structured data.
• Example use cases:
  • Generating product detail pages and product cards.
  • Dynamically generating survey forms for users to fill in; results can be resubmitted to the Agent as structured data.
  • Dynamically generating forms for users to select and fill in; results can be resubmitted to the Agent as structured data, with keywords triggering specific FlowAgent branches.
• Example prompt for an LLM-generated A2UI protocol JSON:

                      
                      {"version": "v0.9", "updateComponents": {"surfaceId": "main", "components": [
  {"id": "header", "component": "Text", "text": "# Book Your Table", "variant": "h2"},
  {"id": "date-picker", "component": "DateTimeInput", "label": "Select Date", "value": {"path": "/reservation/date"}, "enableDate": true},
  {"id": "submit-text", "component": "Text", "text": "Confirm Reservation"},
  {"id": "submit-btn", "component": "Button", "child": "submit-text", "variant": "primary", "action": {"event": {"name": "confirm_booking"}}}
]}}

                    

3.187.0402（2026-04-02）

API

• Added "Update Q&A Knowledge Document" API
• Added "Add Q&A Chunks" API
• Added "Add Q&A Knowledge Document" API

V3.184.0331（2026-03-31）

Integration

• Added segmented output for Line

FlowAgent

• Message Pass-through component optimization: When human service is triggered, messages from the message pass-through component can be sent to end users via API or third-party channels.

API

• Added statistics API for Agent conversation credit consumption statistics

V3.181.0324 (2026-03-24)

Log - AI Quality Check

• Added the AI Quality Check feature to Agent/FlowAgent logs, allowing production conversations to be evaluated against existing Q&A data in the knowledge base.
• AI Quality Check generates an overall score and improvement suggestions at the Conversation ID level, and scores with rationales across 5 dimensions at the Message ID level.
• The 5 evaluation dimensions are: factual_correctness_score, faithfulness_score, relevance_score, fluency_score, and safety_score.

Integration - LiveDesk

• Added a message-sync switch under Agent - Integration - LiveDesk. When enabled, GPTBots automatically syncs messages sent from LiveDesk.
• Agent and user messages from LiveDesk can be used as context for the AI Agent, improving the accuracy and quality of the Agent's responses.

V3.175.0316（2026-03-16）

MCP

• MCP Headers and Query parameters now support variables

SSO

• Added support for OIDC protocol
• Added SSO access token to global variables

3.171.0311 (2026-03-11)

Knowledge Sources

• Attachment messages uploaded during conversations can now be used as knowledge sources for users to review within the conversation. When the knowledge source display mode is set to Tag, knowledge sources are shown as tags within the conversation.
• Optimized the knowledge source preview window: in Tag mode, clicking the citation badge number on a knowledge source loads a preview of the PDF file and jumps to the specific page, highlighting the referenced area. This helps users quickly locate the source and verify the credibility of the information.
• The Message API also adds support for the enhanced knowledge source citation feature. See the Message API documentation for details.

3.157.0203（2026-02-03）

FlowAgent

• Added Tools Node: Supports inserting Tool nodes into FlowAgent to enable direct tool calling. Input parameters must be provided manually or by referencing variables, and execution results can be referenced as variables by downstream nodes.
• Added Workflow Node: Supports inserting Workflow nodes into FlowAgent to enable direct workflow calling. Input parameters must be provided manually or by referencing variables, and execution results can be referenced as variables by downstream nodes.
• Added Variable Assignment Node: Supports assigning values to User Attributes and Custom Attributes. Assignment operations include overwrite, append, and clear. Assignments take effect immediately, and downstream nodes can reference the updated variable values.

Attachments

• Optimized attachment recognition control for FlowAgents with multiple LLM Nodes. LLM parameters now support controlling recognition strategies for different attachment types (images, videos, audio, files), effectively reducing token consumption.

Logtree

• Agents and Workflows now feature the new Logtree logging system. Logtree displays node execution processes chronologically and hierarchically, providing detailed log information to enhance observability and transparency, helping users quickly identify and resolve issues.

ASR

• ASR supports custom prompts and automatically incorporates the previous round of conversation by default to improve recognition accuracy and output language consistency.
• The Language preference list has been optimized to cover more languages.

3.152.0127（2026-01-27）

Agent Integration - LiveDesk

• The LiveDesk channel now supports segmented output, allowing Agent responses to be sent to LiveDesk in segments. This currently includes text, image, video, file, and audio segments.
• The variable feature on the Agent configuration page now supports referencing LiveDesk channel variables: ld_user_id, ld_contact_id, ld_phone, and ld_email.

Agent Integration - Telegram

• When receiving group chat messages from the Telegram channel, the system now supports appending User Name information to messages to facilitate user identification within group chats.

3.148.0121（2026-01-21）

Agent Integration

• Added support for new global LiveDesk variables, including ld_identifier, ld_contact_id, ld_phone, and ld_email. User identity IDs set in LiveDesk—such as those provided through pre-chat forms or via APIs—are automatically synchronized to GPTBots as the user’s identity identifier. Additionally, conversation attributes filled in by users are automatically mapped to GPTBots custom attributes and assigned accordingly.

3.146.0119（2026-01-19）

Tool

• Added support for OAUTH tools, currently integrated tools include: Gmail, Google Docs, Google Sheets, Google Calendar.

V3.145.0114（2025-01-14）

Agent Integration

• Added support for Teams platform, allowing Teams users to chat with Agents in single chats, groups, and channels.