Concierge Agents

Concierge Agents are tenant-scoped AI workers that can answer customers, use knowledge and skills, categorize conversations, and hand off to humans. Coming soon: Concierge configuration, Skills and Optimization Agent in CXP are Owner/Admin-only. Existing Insights and Analyst permissions remain separate, and legacy-app/shared Optimization access is unchanged.

Prerequisites

You need administrator permissions for Concierge Agent management.

Agent directory

Open Agents > Concierge to find and open Concierge Agents. Use the compact directory toolbar to search agents or choose visible columns, then scroll the continuous directory instead of switching between numbered pages. The directory progressively reveals 50 agents at a time as you approach the bottom, and continued scrolling can reach every matching agent. Each row shows the agent name with a one-line description, its runtime status, versions, and last update date; Type and Impl can be turned on from Columns. Select a row to see its description, live and draft versions, type, and last update in the preview rail, then select Open editor at the bottom to enter the editor. Long names are shortened with an ellipsis in the table while the complete name remains available in the agent record.

CXP Concierge Agents directory showing tenant agents, versions, and agent status

Edit agent configuration

The agent editor keeps your location visible while you work. The left navigation identifies the selected agent, highlights the active section, and provides All Concierge Agents at the top to return to the directory. If the draft has unsaved changes, returning asks whether to stay, discard the changes and return, or save the draft and return; a failed save keeps you in the editor. The content header repeats the current section and keeps the dirty-gated Save draft action available beside it. The right rail shows the agent status with Refresh and a More actions menu for Delete agent, followed by a single Versions list. Selecting a version expands its details in place: the draft shows its changes since the current version with Publish draft and Discard draft, and published versions show an inline-editable label, release notes, and Restore to draft. Hover a published version to show its pin button; pinned versions keep a highlighted pin. Publish draft opens a dialog for the required release notes and an optional version label, and is unavailable until unsaved edits are saved. Deleting an agent requires a second, explicit confirmation. Navigation descriptions fit on one line; hover or focus a section to read its full description. Skills follows Instructions, Messages follows Knowledge, and Collect Data appears immediately before Business Hours. Depending on the agent type and version, you can manage:

Instructions and behavior
Requirements for authoring and Optimization
Knowledge references
Knowledge cues
Analysis settings, including category references, CSat, and summaries
Skills
Channel-specific behavior
Customer-facing messages
Customer-data collection
Business hours and closed-hours messages
Voice responder settings when enabled
Draft, current, and historical versions

Some older or unsupported agents may be read-only or require an upgrade flow before editing.

Instructions and advanced settings

Use Instructions for the agent’s main behavior guidance. This is the field previously labeled Personality. Keep reusable behavior in assigned Skills and channel-specific formatting in Channels. The Instructions editor fills the central pane. Long instructions scroll within the editor while the section header and Save draft remain available.

CXP Instructions editor filling the central pane beside compact section navigation and the Versions rail

Advanced groups model and runtime settings, general guardrails, and advanced JSON configuration. Edge voice responder instructions appear when an Edge voice responder model is configured. Clearing the model hides those instructions and preserves their saved text. Preflight Model selects the model used to understand customer intent and prepare knowledge queries. Leave it blank to use the runtime default. Preflight Depth sets how many recent messages it considers. The default is five; a value of zero also uses the runtime default of five. Values must be zero or greater. Force Handoff on Guardrails sets the number of accumulated guardrail events in a conversation that triggers handoff. Zero disables this threshold. Events do not have to be consecutive; preflight rejections do not count toward the threshold. Set fallback wording for automatic handoff under Messages > Default Handoff Message.

CXP Advanced settings with preflight models, voice responder instructions, and the guardrail handoff threshold

Customer-facing messages

Open Messages, immediately after Knowledge, to configure predefined customer-facing text:

Setting	What to configure
Engage message	A proactive invitation beside the closed Chat Widget. Blank uses the default invitation.
Engage timeout	The delay in seconds before the invitation. The default is five seconds; zero also uses five.
Welcome message	The initial greeting. Blank uses the default greeting.
Personalized Voice Welcome message	A voice greeting template, such as `Hello {name}, how can I help?`. It uses available contact fields and falls back to the ordinary greeting when a placeholder cannot be filled.
Disclaimer message	A disclosure shown in chat before the first customer message. Leave blank to hide it.
Decline message	The response to a rejected request. Blank uses the default decline response.
Handoff message	The customer-facing handoff text, or generation guidance when personalization is enabled.
Personalize Handoff Message	Generate handoff wording from the conversation context and the configured handoff guidance. This also applies to closed-hours handoff guidance.
Default Handoff Message	Fallback wording for automatic handoff paths. Blank uses the default text. This field is not personalized.

The engagement invitation is for the closed Chat Widget; it is separate from voice re-engagement messages in Channels. Business Hours and holiday settings can override the welcome and handoff messages. Save the draft and test the relevant conversation flow before publishing.

CXP Messages section with engagement, welcome, disclaimer, decline, and handoff settings for a saved Concierge Agent draft

Channel behavior

Use Channels to adapt the same agent to each medium:

Setting	What to configure
Chat Instructions	Formatting and behavior for chat messages.
Email Instructions	Formatting and behavior for email replies.
Voice instructions	Spoken response style, such as how to read dates or numbers.
Engage Messages for Voice	Alternative prompts to re-engage a silent caller, one per line. Used when the agent’s voice engage timeout is enabled.
TTS parameters	Text-to-speech voice configuration, as a `provider:voice` value or JSON. Leave blank to use the configured default.
ASR parameters	Speech-recognition configuration as JSON. Leave blank to use the configured default.
Enable platform-side audio recording	Whether the platform records call audio. Other recording arrangements remain separate.
Multimodal instructions template	How to combine voice and chat responses. Retain the `{{voice_instructions}}` and `{{narration_delimiter}}` placeholders when customizing it; leave blank for the shared default.

CXP Concierge Agent Channels section with separate chat, email, and voice instructions in a saved draft

Include images controls images in the agent’s answers. Upload files accepts customer files. Enable Analyze uploads to analyze supported images and PDFs before the agent responds. Without analysis, the files remain attached to the conversation. When analysis is enabled, Image Analysis Guidelines can specify what to extract from images and how to describe them. Leaving the guidelines blank uses the default image description and text extraction. Show Image Description to User controls whether that analysis text appears in chat; the agent can still use the analysis when the description is hidden. Turning off Analyze uploads hides the guidelines and description switch while preserving their values.

Collect customer data

Open Collect Data to define the information the agent collects from a customer. Select Add field, then choose a Field type, a reference Field name, a customer-facing Display label, and a Required setting. Required fields are collected before handoff to a person.

Use User name, Email address, or Phone number for standard contact details. Each of these types can appear once. Their reference names are fixed as name, email, and phone.
Use Text for another free-text value or Select for a choice from a list. Give each custom field a unique name without spaces. For Select, enter the choices in Select options, separated by commas in the order they should appear.

Turn on both Collect data at startup and Use a form in chat to show the configured form when chat starts. The startup setting alone does not make the agent ask for the details conversationally. With the form off, the agent can collect details in the conversation when needed for handoff. For agents using a pre-v8 implementation, the editor also shows Collect data message and Collect data guidelines. These settings remain saved when they are hidden for an agent using v8 or later. Save the draft and test the collection flow before publishing changes.

CXP Collect Data section showing configured customer name, required email, and order number fields

Business hours and closed messages

Open Business Hours to configure the weekly schedule and timezone used for open- and closed-hours behavior. This section replaces the former Handoff section. General handoff messages are now in Messages. When Business Hours is enabled, configure these closed-hours messages in order:

Welcome message when closed for the opening greeting.
Personalized voice welcome message when closed for a voice greeting template using contact fields such as {name}.
Handoff message when closed for handoff text or, with personalization enabled in Messages, guidance for generating that text.

Turning off Business Hours hides the closed-hours message fields and preserves their values. Use Holiday windows for dated closures or temporary message overrides. Save and publish schedule changes through the same version workflow as other agent settings.

CXP Business Hours section with a weekday support schedule, timezone, and holiday controls in an unsaved draft

Conversation analysis

Open Analysis, immediately after Business Hours, to select the Categories catalog for conversations and the Handoff categories catalog for handoff reasons. Turn on Enable CSat to configure customer-satisfaction scoring. Use CSat Guidelines for scoring guidance and Relevant Domains to describe the customer-service topics eligible for analysis. Minimum messages before CSat sets the minimum conversation length before a score is generated; it defaults to four and accepts a number of at least one. Customer, assistant, and human-agent messages count toward the threshold. Turning off CSat preserves its settings. Use Summarize Guidelines to specify the content, format, and length of conversation summaries. Leave it blank to retain the default summary behavior.

CXP Concierge Agent Analysis section with categorization and enabled customer satisfaction settings

Knowledge cues

In Knowledge, select the knowledge base and configure citations, retrieval guidance, and knowledge guardrails. Turn on Enable cues when the agent should narrow knowledge retrieval or add conversation tags based on where the widget is placed or on information collected from the customer. Select Add web page cue or Add collected data cue to create an entry.

Web page cues match the widget’s referring page with a regular expression in URL pattern. For example, ^https://www\.example\.com/returns/ matches pages under that site’s returns section.
Collected data cues use a JavaScript Condition that evaluates to true for the collected contact fields. For example, typeof topic !== "undefined" && topic === "returns" matches when the collected topic is returns.

For each cue, choose a Knowledge filter of None, KB document, or KB tags. A document filter uses its exact KB document source; a tag filter takes comma-separated Knowledge tags. Conversation tags is a separate comma-separated list of tags to add when the cue matches. Web page cues run before collected data cues. Later matching cues replace the earlier knowledge filter, including a matching None filter that removes the earlier restriction. Conversation tags from matching cues accumulate. Turning off cues preserves the entries so you can enable them again later. Correct invalid URL patterns before saving the draft.

CXP Knowledge Cues section with web page and collected customer data rules, knowledge filters, and conversation tags

Manage Requirements

Requirements describe the business rules and expected outcomes that guide agent authoring and Optimization. You can reuse one Requirement across several Concierge Agents. Requirements do not directly add instructions to customer conversations; use the agent’s Instructions or Skills to implement the behavior.

Create or edit a Requirement

Open Requirements under ADMIN.
Select Create requirement, or search the directory and open an existing one.
Enter its ID, name, description, and Markdown under Requirement content. For example, a Returns policy Requirement can describe eligibility, required order details, and when a human review is needed.
Select Save.

The ID is fixed after the first save so existing agent associations remain stable. Saved Requirement content is shared by every agent that references it; Requirements do not have a separate draft or publish step. Use Import in the directory or Export in the editor to reuse Requirements as JSON. Importing an existing ID asks you to confirm replacing its shared content.

CXP Requirement editor showing a stable ID, descriptive metadata and Markdown customer support standards

Associate Requirements with an agent

Open the Concierge Agent and select its draft.
Select Requirements, immediately before Instructions in the editor.
Select Add requirement and choose a Requirement that describes the agent’s intended behavior. Repeat to add more. Use Change or Remove on an assigned card to update a selection. Manage requirements opens the shared directory in another tab so your agent draft stays open.
Select Save draft, then publish the agent’s association changes through the usual version workflow.

Removing an association leaves the shared Requirement available to other agents. If a referenced Requirement is unavailable, its saved reference stays visible so you can replace or remove it. Before deleting a shared Requirement, review the agents that use it.

Support Concierge draft with two assigned Requirements before the Instructions section

Test an agent version

Open an agent and select Test to try its customer Chat Widget in your current organization. Choose Draft, Current, or a historical version in the Versions rail. The chat uses that version’s saved configuration; unsaved edits are not included. The notice beside the chat identifies the version being tested. An agent must have an initial published version before you can use Test, including when testing a draft or historical version. Once that prerequisite is met, you can test a saved draft without publishing its changes. Widget sign-in continues to use the current published authentication settings when you test another version.

CXP Concierge Agent Test page with the saved draft Support Concierge greeting and a New conversation action

Select the version you want to test. If you are editing a draft, select Save draft to include your latest changes. Saving the selected draft while Test is open resets the widget with the saved configuration.
Wait for the widget to load and review the configured welcome message or opening form, if the agent has one.
Send a representative customer question. Check the answer, follow-up questions, and any configured handoff or integration behavior.
Select New conversation to reset the widget for another scenario, or select another version to compare behavior in a fresh browser session.
Leave Test when you finish. Returning to Test or opening another agent starts a fresh browser session.

Test uses the real agent and its configured tools. Use synthetic customer details for your scenarios. The test session’s browser storage is temporary; starting over does not delete conversations already recorded in CXP. Your app session and other test tabs retain their own state. If the widget cannot load, select Try again after checking the connection. To supply test visitor details, expand Context data, enter a JSON object, and select Apply context. For example:

{
  "email": "visitor@example.com"
}

Applied context is reused by New conversation. Leaving Test or switching agents clears it. An empty field applies an empty object. Normal contact matching still applies: supplying an email can associate the test with an existing contact and live conversation.

Holiday windows — Coming soon

In Business Hours, use Holiday windows to prepare dated closures or message notices in the agent’s Business Hours timezone. Apply and preview the draft, then save and publish through the existing controls. The window expires automatically. See Holiday windows for setup, message behavior, and copying selected dates to next year.

Optimize an agent

Open Agents > Optimization Agent, select a Concierge Agent, and describe the change you want. Explain this Agent fills in a configuration-review prompt; Verify this Agent fills in a prompt to run evals or create representative evals when none exist. Review the prompt, then send it.

CXP Optimization Agent start page with the agent selector, reasoning slider, and example prompts

Use the slider below the prompt to select Medium, High, or Extra High reasoning. High is the default for a new session. You can use arrow keys, Home, or End when the slider has keyboard focus. In an existing session, the slider starts at the last submitted supported level. Changing it while work is running prepares the next message; it does not change the current work. Open a session from the rail to review progress or send follow-up requests. Stop requests that the running turn stop. Expand the rail when needed to switch sessions or reload a session’s history. The session header lets you copy its link or download the full trace for troubleshooting. If a request times out, inspect the session before retrying because the original request may have been received. Administrators can open Settings > Optimization Agent to review instructions and their built-in, system, or tenant source. Reload refreshes the installed built-in files and configured overrides. Review the resulting agent draft and eval results before publishing a version.

Targeted instruction edits — Coming soon

The Optimization Agent will be able to change a selected passage in Personality, Skills or Requirements without regenerating the surrounding instructions. It will check that the text still matches the version it read and report the committed differences. Repeated wording will need enough surrounding context to identify the intended occurrence. If the field changes in the meantime, the agent will need to read it again before applying the edit. To retain control of configuration changes, give an explicit advisory instruction:

Do not edit configuration, Skills or Requirements yet.
Propose exact changes and ask for approval to apply them.
You may create and run targeted evals for this task.

When you are ready, approve the proposed edit. For example, “Apply the proposed email clarification bullet to the draft; preserve all other text.” The agent should apply the approved change. Review the reported differences separately from eval results: passing email evals does not show whether an unrelated product list was preserved. Skills and Requirements can be shared by multiple Concierge Agents, so review their impact before approving those edits. The agent will also read relevant sections of long instructions and sources, check whether previously read text has changed, and reuse retrieved evidence. Conversation reviews will avoid repeating the same transcript, and eval searches will focus on the relevant tests. Complete source text will remain available when needed. When you request a model migration, the agent will review Personality and Skills against the target model’s prompting guidance and make justified adaptations while preserving intended behavior. It will explain any decision to keep the wording unchanged and report validation gaps without weakening existing eval requirements. Advisory wording guides the agent’s behavior; it is not a separate access-control mode in the interface. This update is not yet confirmed available in production. Read-only reviews will use the published version when no draft exists and will leave draft creation until a configuration or eval edit. Requests to review or show proposed changes will remain advisory; clear approval of a concrete edit will let the agent proceed without another confirmation. For new or changed privacy and authorization evals, the agent will inspect the first passing conversation before reporting validation. If the conversation contradicts the judge, it will explain the discrepancy. Mismatched results will lead to a focused check or a justified retry of the same eval, with unresolved validation reported explicitly. Created evals, including diagnostic probes, will be included in artifact counts. Fields unavailable through optimization tools will be identified for manual configuration. Reports will distinguish changes saved in the current turn from earlier work, unchanged saves, and proposed changes. Reusing fewer evals will include checking that important scenarios still have both suitable inputs and required assertions. Saved test coverage will be reported separately from successfully validated behavior. Focused evidence reads and clearer edit diagnostics will help the agent recover without repeatedly loading entire conversations or rewriting instructions.

Handoff behavior evals — Coming soon

Handoff behavior evals will count a call to handoff_conversation as completion of the handoff step. A promise to connect the customer will not be sufficient. Required contact details, appropriate timing, and message requirements will still be checked. Downstream email delivery, routing errors, or whether a human has joined will not fail that step; validate those outcomes separately through integration testing. A passing behavior eval will not prove successful delivery.

Manage versions

CXP uses draft and published agent versions so administrators can review changes before they affect customer conversations. Typical actions include:

Create or edit a draft.
Save draft changes.
Publish a version with release notes.
Restore prior behavior to a draft.
Pin or label versions when supported.

Review the draft’s instructions and eval results before publishing. The editor’s Test section lets you check the saved draft before publishing its changes, provided the agent already has an initial published version.

Recoverable agent deletion — Coming soon

In the Versions rail, open More agent actions and choose Delete agent. This will move an agent to Deleted Assistants, preserving its drafts, published versions, history, and evaluations. The agent will disappear from normal CXP lists and selectors. A system administrator will be able to restore the same agent from Deleted Assistants in the legacy application. Tenant administrator access alone will not grant access to that recovery page. Agents permanently removed before this change will not become recoverable. To discard edits to an already published agent, select its draft and choose Discard draft in its expanded version entry. A separate confirmation will warn that the discarded draft cannot be restored. To publish saved edits, choose Publish draft in that card and enter release notes in the dialog. This removes the working copy while keeping the published version. A new agent with only its first draft has no published version to retain, so it will offer Delete agent instead. Before deleting a channel-assigned agent, change the channel configuration as needed. Moving an agent to Deleted Assistants will not automatically clear existing channel assignments or stop active conversations or Optimization work.

Skills

Skills extend agent behavior. Open Skills under ADMIN to manage tenant skills, then assign them to Concierge Agents with the appropriate load mode. For Agent Assist, assign the Agent Assist skill only when you want that agent to produce internal operational guidance and reviewable action suggestions.

Automatic conversation tagging

Define when an agent should add conversation labels and how current tags should guide its responses. Keep simple rules in the agent’s instructions, or assign a shared Tagging Policy skill to several agents. See Automatic conversation tagging for availability, copyable examples, skill setup, and verification steps.

Channel assignment

Concierge Agents become customer-facing when assigned to channel entry points such as chat, email, or phone settings. Changing a channel’s assigned agent affects future activity for that entry point. Before changing a production channel, confirm the agent has the right instructions, knowledge, handoff behavior, and category configuration.

Getting started

Agent guide

Admin guide

Scenarios

Reference

Concierge Agents

Concierge Agents

Prerequisites

Agent directory