Release Notes

Version 1.14

New Features

• One agent can now consult another. An agent can ask a different agent a question in the middle of a conversation and use the answer. A general assistant that runs into a specialist question can consult the agent that owns that knowledge, instead of guessing or sending the person elsewhere — and the consulted agent answers using its own instructions, its own tools and its own data access. It runs as the person using the assistant, so it can only see what they could already see, and it cannot change anything on its own: anything it proposes comes back into the original conversation for that person to confirm. Each consultation costs one extra AI turn and is kept as its own conversation, so the exchange is auditable. Nothing happens until you switch it on. On the agent that should be available to answer, tick Agent Invocable and Allow Automated Requests — both are required, and if only one is ticked the calling agent gets a clear message naming the one that is missing. On the calling agent, Max Agent Call Depth controls how far consultations may go: 1 (the default) lets it consult one agent, 0 means never. One thing to know: the consultation appears in the transcript through the calling agent’s own words and a tool indicator — if you want people to see which agent was consulted, tell the calling agent to say so in its instructions. See Configuring Agent Settings and Built-in Tools Reference.
• Choose who an email comes from, and where replies go. Until now an agent sending email had no say in the From address: it used whichever organisation-wide address the system returned first or, if your org had none, the account the automation runs under — a service account rather than a person or a team. Nothing set a reply address either, so a recipient’s reply had no defined route back. An agent’s instructions can now name the sending identity (an organisation-wide address that already exists and is verified), the display name the recipient sees, and the reply address. That last one is what closes the loop: point it at an address wired to inbound routing and the reply comes back into the conversation instead of landing nowhere. If the sending address you name cannot be matched, the email is still sent using the default sender and the mismatch is recorded — an agent never goes quiet because someone mistyped an address. Nothing changes unless you ask for it: an agent whose instructions mention none of this behaves exactly as it does today. See Outbound Email and Email Reply Handling.

Improvements

• Dark mode everywhere. The chat widget your customers see, and every REDA AI admin screen — the console, prompt builder, tool builder, conversation history, analytics, transcript, incident debugger, email routing and the rest — now follow your org’s Salesforce theme, including dark mode. They were previously fixed to a light palette whatever theme you were using. Two things stay deliberately fixed because they read better that way in both modes: dark code-preview panels, and the colour-coded status badges. There is nothing to turn on.

Bug Fixes

• Scheduled-action conditions now reliably stop an action. When an agent schedules something for later it can attach a condition that is checked again at the moment it fires — “only send the follow-up if they still haven’t replied”. The condition exists to stop the action. If it could not be checked — an unsupported comparison, or a field or object that had been renamed or deleted — the action ran anyway; and actions triggered by an email reply ignored their condition entirely. In both cases the guard was quietly doing nothing. Now, a condition that cannot be checked skips the action and records the reason on it, so it is visible rather than silent; a check that fails because of a temporary error is retried and only reported if it keeps failing; and email-reply actions honour their conditions like any other. The comparisons available are equals, not_equals, contains, not_contains, is_blank and not_blank — typed exactly like that, in every language — with contains and not_contains new in this release. See Scheduled Actions.
• Assistant text was unreadable on the newer Salesforce theme. On orgs using the newer Lightning styling, the assistant’s message box showed white text on a white background — you could not read what you were typing — and the confirmation cards for emails and tool results were unreadable for the same reason. Text now uses colours that adapt to the theme, in both light and dark. Orgs on the older theme were never affected.

Version 1.13

New Features

• A docked side panel for the Community & Website assistants. The assistant can now open as a full-height panel docked to the right or left edge of the page, instead of a floating card above the launcher. It gives answers much more room to breathe — property cards, photo sliders and longer replies no longer feel cramped — and the page stays readable alongside it, so visitors can keep browsing while they chat. The launcher bubble is unchanged and still opens and closes the panel, and on phones the panel opens full-screen exactly as before. On the Community portal, pick Display Mode (Bubble, Docked Right, Docked Left) on the REDA AI component in Experience Builder. On a website, add data-display-mode="docked-right" (or docked-left) to your embed, plus an optional data-dock-width if you want to set the width yourself instead of letting it adapt to the screen. Existing widgets are unaffected — leave Display Mode alone and everything looks and behaves exactly as it does today. See Widget UI Customization → Display Mode and Community Channel.
• The Community assistant now matches your portal’s colors automatically. Leave Primary Color blank and the widget picks up your community’s own brand color from your Experience Builder theme — the launcher bubble, header and message bubbles all follow it, so the assistant looks like part of your site with no setup at all. Enter a hex value only when you want to override it. Communities that already have a Primary Color set keep exactly what they have. See Community Channel → Branding.

Improvements

• More reliable credit tracking when many people use your agents at the same time. When a large number of conversations happened to run at the very same moment, credit updates could briefly collide and leave a conversation waiting for its reply. Credit tracking has been reworked to handle heavy concurrent use smoothly, so replies keep flowing during busy periods and every conversation’s credit usage is still counted accurately. This happens automatically — there is nothing to turn on or configure. See How Credits Work.
• The in-Salesforce assistant now follows your Salesforce theme. The assistant’s backgrounds, text, borders and cards now take their colors from your org’s Salesforce theme instead of fixed values, so it stays visually consistent with the rest of Salesforce — including dark mode, where the assistant now adapts instead of staying stubbornly light. Your own branding still wins: any color you set on the component is used exactly as given. The difference is what blank means — Header Color, Header Icon Color and Chat Background no longer carry a fixed default, so leaving one blank now means “follow the Salesforce theme”. In light-themed orgs the assistant looks the same as it did before. See Internal Channel.
• Accessibility improvements. The photo slider on cards can now be operated from the keyboard — tab to the arrows and press Enter or Space to move between photos — and the assistant’s replies are now announced to screen readers as they arrive, instead of appearing silently.

Bug Fixes

• Allowed Origins now reliably restricts which websites can use your Website widget. When Allowed Origins was filled in on an agent, the restriction was not applied reliably, so a key copied from your page could still be used from another website. REDA AI now checks the site each request genuinely came from, as reported by the visitor’s browser, which a page cannot fake. If Allowed Origins is blank — the default — nothing changes for you and every origin is still allowed. Two notes if you do use it: list every domain your visitors reach you on (both example.com and www.example.com, if both serve your site), and if you call the chat API directly from your own server rather than from a browser, leave Allowed Origins blank and control access another way. See External Widget → Configuring Allowed Origins.

Version 1.12

New Features

• Branded email. Agent emails can now carry your brand — logo, colors, header and footer — instead of plain text. Store your brand once in a new Branding record and point an agent at it with the agent’s Branding lookup; outbound email then sends a branded HTML message (plus a plain-text part) with the sender’s signature above your footer. It’s opt-in per agent with no org-wide default — an agent with no brand linked keeps sending plain text — and works across the Internal, Widget, Community and API channels. A new Optimize Branding with AI panel reads an uploaded brand document (letterhead, guidelines, brochure) and proposes the branding fields for you to review, in the document’s own language. See Branded Email.
• AI Recommendations on record pages. A new REDA AI Recommendations component (Lightning App Builder) shows agent-generated next actions and risks for the record in view — a short takeaway up front, full detail in a “View details” window, with 👍/👎 feedback. Advisory and read-only, works on any standard or custom object, responds in the user’s language, and lets you toggle which sources (emails, tasks, events, files, notes, field history) it scans. See AI Recommendations.
• Deep Reasoning per agent. A new Deep Reasoning setting lets you choose how thoroughly an agent works through a question before it answers — from Off (fast, direct answers) through Low, Medium, High, Xhigh to Max (the most careful, step-by-step reasoning). Higher levels improve accuracy on complex, multi-step questions, at the cost of a slower reply and more credits; Off is the default and is the right choice for most everyday agents. The level applies to everything that agent does — chat, actions and AI Recommendations — so you tune reasoning depth once, per agent. See Deep Reasoning.
• Inbound email without attachments now reaches your agent. The Inbound Email channel now hands any email at a routed address to its agent as a conversation — the body, subject and sender become the agent’s message, with or without attachments. Previously a body-only email was accepted but silently dropped; only emails carrying a qualifying attachment reached the agent. Turning on the channel is just the routing setup (an Email Service address plus an active routing record) — there’s no separate switch. See Email Channel — Inbound.
• Voice input on the Community & Website widgets. Portal users and website visitors can now dictate messages by voice using a microphone button; their words appear in the message box as they speak, to review before sending. On by default where the browser supports it — turn it off with Enable Voice Input (Community, in Experience Builder) or data-enable-voice="false" (Website embed). See Voice Input (Speech-to-Text).
• Display modes for the in-Salesforce assistant. Choose how the assistant appears on a record, app, or home page — Embedded (inline), Docked Right/Left (a side panel that slides in over the page), or Chat Bubble (a floating button that expands into a panel). Docked and Chat Bubble open full-screen on phones and the Salesforce mobile app. See Internal Channel.
• Multi-step autonomous workflows for API & Flow agents. An API-channel agent can now carry a task across several automatic turns — for example, create a record, call an external service using that new record, then write the result back. Each step runs on its own, so an agent is no longer limited to a single outside call before it saves data; the agent decides when another step is needed. Available only on the API (Flow / programmatic) channel and on for API agents by default — remove it from an agent’s loadout to opt out — with a safety limit on how many steps one task can chain. See Invoking REDA AI from Flows.
• File attachments on the Community portal widget. Signed-in portal users can now attach files to their messages using a paperclip button — up to 5 per message, each shown as a named chip on the sent message. Uploaded files are automatically linked to the customer’s Contact record, and the assistant can attach a file to a related record (such as a Case it raises) when asked. Appears only when the agent has Allow File Upload enabled. See File Upload and Community Channel.
• Multiple files via Flow & API. The REDA AI invocable action now accepts several files in a single call, so an automated flow can send multiple documents to the assistant at once. See File Upload.
• Photos & video in cards. Record cards can now include a photo slider and video / virtual-tour tiles. A card with several images becomes a swipeable gallery — arrow buttons on a computer, swipe on touch screens and the Salesforce mobile app — with a photo counter; a video or tour shows a play tile that opens in a new tab. Ideal for property and unit listings, product results, or any record with visuals, and it works on the internal Salesforce chat, Community, and website widgets. See Card-List Rendering → Photos & Video.

Improvements

• “Email File Processing” clarified as an attachment control. The Email_File_Processing_Enabled setting governs only whether file attachments on inbound emails are processed — it no longer gates the channel itself. With it off, a routed email still reaches the agent as a conversation; its attachments are simply ignored. See Email Channel — Inbound.
• HTML-only emails are read correctly. When an inbound email has no plain-text part, the platform now passes the stripped HTML body to the agent instead of an empty message.
• Modernised in-Salesforce assistant. A lighter, cleaner header and welcome screen, more readable typography and spacing, smoother open animations, a “scroll to latest” button, and a live status while the assistant is working. Colours, fonts, and the header remain fully configurable in the Lightning App Builder. See Internal Channel.
• Voice dictation is now inline (Internal). Speaking to the in-Salesforce assistant streams your words straight into the message box instead of opening a separate dialog. See Voice Input (Speech-to-Text).
• Clearer, consistent answer formatting. Section headings, lists, and tables now render cleanly and with even spacing across the in-Salesforce assistant, the community widget, and the website widget.
• Attach up to 5 files per message. The in-Salesforce assistant now accepts multiple attachments on a single message (previously one). Each shows as its own named chip and can be removed individually; the agent processes them all together. See File Upload.

May 2026

New Features

• Conversational actions on Community & External channels. The agent now executes a confirmed action automatically and replies with the real outcome (e.g. “I’ve created Case #00001234”) instead of showing a confirmation card; the Internal channel keeps cards. Capped by the new Max_Continuation_Depth setting. See Action Confirmations → How confirmations differ by channel.
• Custom Tool security acknowledgment. Creating a client-defined custom tool records a shared-responsibility acknowledgment (your name + timestamp). See Apex Class Tools.
• Attach a sample file when optimising. The Prompt Optimizer and Build Agent with AI accept a sample document (PDF, image, CSV, TXT, JSON) plus up to 15,000 characters of instructions. See Prompt Optimizer and Build Agent with AI.

Improvements

• Smarter conversation memory (never truncates). History is now a token-budget window of complete messages — no mid-message truncation. Tune it with the new History_Token_Budget setting (default 30,000); five older history settings were retired. See Platform Settings and the Settings Reference.
• Accurate aggregate queries. Grouped/aggregate (GROUP BY) tool queries are no longer capped by the safety row limit.
• Consistent tool-failure reporting. When a tool fails, the agent reports the real reason instead of implying success.
• Refreshed default prompts. The conversational action flow ships with updated default prompts. Orgs that previously seeded the defaults can re-sync by re-running Seed Default Data in the Admin tab — it upserts the latest default Agent Definition, starter prompts, and suggestion chips (existing records with matching names are overwritten). See Platform Settings.

Bug Fixes

• Configuration Health no longer raises a false “Context Fields never auto-load” warning for API, Community, or External agents.
• Unreadable stored tool context now logs an incident instead of failing silently.
• Tool Builder “Deactivate” works from the list view and refreshes on completion.
• The AI-generation spinner stays centred in view instead of appearing below the fold.
• Fixed mid-response truncation that could corrupt long tool results.
• Standardised modal close buttons for consistent rendering across the app.