AI Agent Builder 8.2.2

What's New

AI Agent Builder

• Real-Time Token Streaming: AI Agent responses can now stream token-by-token directly into the Preview screen and the runtime chat UI, giving users immediate visual feedback as the agent generates its reply. Token streaming can be toggled on or off per agent from the Agent Builder settings panel.
• Guardrails for Input and Output Safety: A new Guardrails layer can now be configured on any agent to enforce safety, compliance, and policy controls. Guardrails can be applied before a prompt is sent to the LLM (input stage) and before a response is returned to the user (output stage). Three guard types are supported: Rule-Based, LLM Safety Check, and Webhook with configurable violation actions including Block, Sanitize, and Redirect. See Guards for more information.
• Context Window Management: Agents can now automatically manage conversation history length to prevent errors caused by exceeding the LLM's context limit. Administrators can configure limits by character count or message count, choose between Truncate or Sliding Window reduction strategies, and control whether system or user messages are preserved during reduction.
• Parallel Tool Execution: When an LLM requests multiple tools in a single response, the agent runtime can now execute those tool calls concurrently rather than one at a time, significantly reducing response latency for multi-tool tasks. The number of parallel threads is configurable.
• Per-Task Timeout Controls: Individual tasks within an agent can now be assigned their own timeout limits, preventing a single slow or stalled LLM call from blocking the entire agent execution. Both agent-level and task-level timeout values are supported, with socket timeout override capability.
• LLM Call Retry with Exponential Backoff: The agent runtime now automatically retries failed LLM API calls caused by transient errors such as rate limits (HTTP 429) or server errors (HTTP 5xx). Retry behaviour is configurable, including initial delay, maximum delay cap, and maximum number of retry attempts.
• Task Decision Making — Skip, Repeat, or Abort: Agents can now be configured to make intelligent decisions after each task completes, choosing to continue to the next task, jump to a specific task, or abort execution entirely. Decisions can be based on Response Condition rules (contains, not contains, equals, and more), a custom BeanShell script, or an LLM-evaluated yes/no judgment.
• Abort a Running Agent: Users can now stop an agent that is currently running, both from the Preview panel during development and from the Agent Execution Audit Trail page. The execution status is immediately updated to ABORTED and reflected in the audit trail.
• User Interaction Tools: Agents can now pause mid-execution to collect input from the user before proceeding. Four interaction modes are available Ask Question, Confirmation, Action Choice, and Get Permission, along with an HTML UI Tool for rendering custom input forms directly in the chat window.
• Rich Display Tools: Agents now have five dedicated display tools for presenting formatted information to users in the chat window: HTML Markup, AJAX-loaded content, URL/iframe embeds, File Viewer, and Script Runner. Content can be shown in a slide-out Display Panel, inline in the chat bubble, as a link, or opened in a new tab.
• Live Streaming of Task and Tool Activity to Chat UI: The AI Assistant chat interface can now display real-time progress updates as the agent works, including task descriptions (e.g., "Gathering issue information"), tool call descriptions, and intermediate LLM messages. Each stream type (Response, LLM Messages, Task Description, Tool Description) can be enabled or disabled independently per agent.
• Agent Variables (Scratch Pad): Agents can now declare named variables that persist across tasks within a single execution. These variables can be set by one task and read by a later task, enabling agents to carry forward intermediate results, counters, or context without relying on the LLM's memory alone.
• Enable/Disable Individual Prompts, Tools, and Enhancers: Each prompt, tool, and enhancer component in the Agent Builder can now be individually enabled or disabled using a toggle switch on its label bar. Disabled components are visually dimmed and are excluded from execution, making it much faster to test and iterate on agent configurations without having to delete, re-add, and reconfigure components.
  
• Advanced LLM Sampling Parameters OpenAI, Anthropic, Gemini, and IBM: LLM provider configurations in the Agent Builder now support optional sampling parameters Temperature, Top P, and Top K (where supported by the provider). Parameters include range validation, and unsupported options (such as Top K for OpenAI and IBM) are automatically hidden.
• AI-Powered Prompt Writer Helper: A new "Write with AI" assistant is available inside the Agent Builder to help users compose high-quality prompts. Users can describe their goal, and the assistant, powered by the agent's own configured LLM, will generate a suitable prompt, which can be inserted directly into the prompt field with a single click. See Text Prompt for more details.
  

AI Composer

• Dedicated Column Formatter Tools for Datalist Composer: The Datalist Composer now uses 12 dedicated tools for applying specific column formatter types (Date, File Link, Font, Image, JavaScript Condition, and Option with multiple binder variants) instead of a single generic tool. This eliminates incorrect formatter definitions that previously occurred when the LLM had to guess class names.

AI Agent Execution Audit Trail

• Token Usage Displayed in Audit Trail: The Agent Execution Audit Trail now shows Prompt Tokens and Completion Tokens for each execution, giving administrators clear visibility into LLM usage costs. The Agent Type column identifies whether an execution originated from AI Composer or AI Agent Builder, and the list now supports date range filtering and sorting by start time.

Bug Fixes

• MCP Client Tool — Nested JSON Arguments Serialised Incorrectly: Arguments containing nested JSON objects (e.g., filter objects) were being corrupted before being sent to the MCP server. This caused MCP tool calls with complex parameters to fail or return unexpected results.
• AI Designer Fails to Generate Designs with Claude Sonnet 4.6 and GPT-4.1: The AI Designer could not complete app generation when using newer AI models. Process designs were not created with GPT-4.1, and form designs were skipped with Claude Sonnet 4.6. All models, including GPT-4.1, Sonnet 4.6, and Opus 4.6, now work correctly.
• AI Credential API Keys Security: Implemented at-rest encryption for API keys configured in the AI Agent Builder to ensure improved data protection and security.
• Datalist Composer Date — Column Formatter Not Working: When asking the Datalist Composer to format a date column as time only, date only, or date and time, the column either displayed the raw unformatted value or failed to parse the date. Date formatting now works correctly for GPT-4o and Claude Sonnet 4.6.
• AI Composer Form Builder Returns Error on Conversational Messages: Sending a casual or generic message (e.g., "Hello") to the Form Builder caused an "Error: Empty response" instead of a friendly conversational reply. All composer builders (Form, Datalist, Process, UI, Theme) now respond naturally to non-design messages.
• Datalist Composer Creates Columns That Do Not Exist: The Datalist Composer was inventing column names that did not correspond to any real fields in the selected form or data source. The composer now validates columns against actual field IDs and will only create columns that exist in the source data.
• "Write with AI" Button Permanently Visible and Blocking Text: The Prompt Writer Helper button was always visible in the Agent Builder, overlapping and obscuring the text content of the prompt and tool property fields. The button now only appears when the user hovers over the relevant field.
• Hint Tooltips Overlapping Admin Bar in Joget DX 9: Guided hint tooltips (Steps 5 and 8) in the AI Agent Builder preview were overlapping the DX 9 admin bar on both Cloud and Local installations, making them difficult to read. Tooltips now display correctly without overlapping.
• AI Composer Carries Over Data When Switching Between Forms: When switching from one form to another in the AI Composer, chat history and session data from the previous form were incorrectly retained and applied to the newly selected form. The composer now automatically starts a fresh session whenever a different form is selected.
• Audit Trail Page Ignores Dark Theme in Joget DX 9: The Agent Execution Audit Trail page displayed a white background when the Joget DX 9 dark theme was active. The page now correctly inherits the active theme's background colour.
• Multiple MCP Tool Reliability Issues: Several MCP tool problems were resolved: repeated queries to the same MCP server were failing with a "Already connected" error; parameter types such as integers, booleans, and nested objects were not being mapped correctly; and enum value validation was incorrectly case-sensitive, causing tools like CoinGecko to fail. All these issues have been corrected.
• Tool Call Recursion Had No Depth Limit: Agent tool calls could recurse indefinitely, potentially causing system instability. A configurable maximum recursion depth is now enforced, and when the limit is reached, the agent gracefully falls back to the LLM's own knowledge rather than failing with an error.

Improvements

• AI Agent Builder Preview Now Shows Guidance When No Node Is Selected: The Preview screen and Agent Execution Detail viewer previously showed a blank panel until the user manually clicked a node. Both panels now display a helpful guidance message by default, making it clear what action is needed to view request and response details.
• AI Composer Redirects to Configuration When No AI Service Is Active: When no AI service is configured or enabled, the AI Composer previously returned a generic error. It now displays a clear prompt with a button that navigates the user directly to the central AI configuration page to set up a service.
• Significant Reduction in LLM Token Costs Across All Composers: The AI Composer agents (Form, Datalist, Process, Theme, and UI/Userview) have been optimised to send substantially fewer tokens on each request. Complete artifact JSON is no longer returned on every tool call, and app context is only included in prompts where it is actually needed, resulting in meaningfully lower API costs.
• Unified LLM Execution Architecture: The internal LLM execution layer has been consolidated so that AI Designer, AI Composer, and AI Agent Builder all route through the same unified engine. This ensures consistent token tracking across all AI features in the Usage Dashboard and simplifies future LLM provider integrations.
• AI Composer Replies Now Use Proposed-Action Wording: The Form Composer's chat responses have been updated to use present tense and proposed-action language (e.g., "I will add a field" rather than "I have added a field"), so users understand that a change is being proposed rather than already applied.
• Redesigned Agent Node Viewer: The node viewer in the Agent Execution Audit Trail and Preview screens has been redesigned with improved visual styling. Nodes now display colour-coded status indicators (green for success, orange for in-progress, red for failed), handle long task names without overlapping adjacent nodes, and can be repositioned by dragging.
• Configurable Maximum Tool Call Depth: In addition to preventing unbounded recursion (see Bug Fixes), administrators can now explicitly set the maximum tool call depth per agent to control how deeply nested tool invocations are permitted during execution.
• Date Range Filtering and Sorting Added to Audit Trail: Alongside the new token usage columns, the Agent Execution Audit Trail list now includes a date range filter (From/To date picker) and supports sorting by execution start time, making it easier to find and review specific executions.

Known Issues

• Date Column Formatter GPT-4.1 Compatibility: The Datalist Composer date column formatter fix is confirmed to work with GPT-4.0 and Claude Sonnet 4.6. A known issue remains where GPT-4.1 may still route to the incorrect tool for date formatting. This is being tracked for a future release.
• User Interaction Tool Default Response Behaviour: When a user interaction tool times out and falls back to a default response, the behaviour may not work as expected in all scenarios. This is being tracked separately and will be addressed in an upcoming release.
• Proposed-Action Wording — Datalist, Process, UI, and Theme Composers: The improvement to use present tense and proposed-action wording in composer replies has been applied to the Form Composer only in this release. The remaining composers (Datalist, Process, UI, and Theme) will be updated in a subsequent release.