AI Agents

Overview

Creating and Managing Agents

Create an Agent

The Agent Card Menu

• Duplicate and Export – see Duplicate and Export below.
• Assign phone number / Assign another number, and Unassign phone number / Unassign phone numbers (N) – see Phone Numbers on an Agent. The unassign item appears only when the agent has numbers.

Delete an Agent

• Phone Numbers In Use – an agent that still has phone numbers assigned cannot be deleted. The dialog lists the numbers; unassign them first. See Phone Numbers.
• Open Email Sessions – if the agent still has open email conversations, the dialog says how many and that they must be completed before the agent is deleted. Resolve & Delete ends those conversations, marks them complete and then deletes the agent. Select Cancel to keep the agent as it is.

The Agent Header

• The agent’s name. Select the pencil icon beside it to rename the agent, type the new name, then press Enter or select the check mark to save. Escape, or the cross icon, discards the change. Studio confirms a rename with Assistant name updated. On a published agent, live conversations keep the published name until your next Publish.
• Copy ID – copies the agent’s ID to your clipboard and confirms with Assistant ID copied. Use it when a setup step asks for the agent ID, such as the SIP header in SIP Integration.
• Assigned number / Assigned numbers – the phone numbers answering for this agent. See Phone Numbers on an Agent.
• Assigned Mailboxes – shown when the agent has Freshdesk mailboxes ticked: the first two, then +N for the rest, with the full list on hover. To change them without opening the agent, select the mailbox chip on its row in Build → AI Agents (see Creating and Managing Agents). See Email Agents.
• Versions – opens Version History, where you see, compare and manage the versions that each Publish saves. It is shown even when there is nothing to publish, and it also appears on Knowledge Base and Data Transformers when you open them from an agent. In a narrow window it shows as an icon.
• Publish – beside Talk and Chat, and shown only while there is something to publish. It reads Ready to publish on an agent that has never been published and Unpublished edits on a published agent whose edits aren’t live yet, and it goes away once everything is published. In a narrow window it gives up its words for a dot and a rocket icon. See Publishing Your Agent.
• Talk and Chat – test the agent. See Testing Your Agent.

Greeting, Voice & Guidelines

Choosing a Voice

Step 1: Selected Voice

• Search voices – type part of a voice name.
• Accent, Gender and Provider – narrow the list. Each filter has a clear (×) button.
• Your search and filters clear when you close the panel or move to Tone + energy, so Selected Voice starts from the full list each time you come back to it.
• Recommended Voices for Multi Lingual – a block of suggested voices at the top of the list. It appears only while the agent is Multilingual with at least one additional language. See Languages & localization.
• Each row shows the voice’s name with its language, gender, accent and provider. A ⚡ beside the name marks a Flash voice, one that is optimized for speed.
• Google’s Chirp 3 HD voices are in the list too, in US English only. Each is listed by its name alone, such as Charon, with Google as its provider and a ⚡ beside the name. Provider has no Google option, so to find one, scroll the list or type its name in Search voices.
• Select the play button on a row to hear that voice. The preview reads the agent’s own Greeting message, so you hear how your real greeting sounds. While the greeting is empty, it plays a short sample line instead, in Arabic for Arabic voices.

Step 2: Tone + energy

Writing Guidelines

• Slash menu. Type / to open a menu of blocks and formatting: Heading 1, Heading 2, Heading 3, Paragraph, Inline Code, Underline, Highlight, Bullet List, Numbered List, Blockquote, Code Block, Divider and Table. Use Search options to narrow the list. Table inserts a 3 × 3 table with a header row.
• Variables. Type {{ to insert a variable. See Built-in variables for what the picker offers.
• Selection toolbar. Select some text to show a toolbar with a block-type menu (Paragraph, Heading 1 to Heading 3, Bullet List, Numbered List, Blockquote, Code Block) and buttons for Bold, Italic, Underline, Strikethrough, Inline Code, Link and Highlight. Link opens a box where you paste a URL, with Apply link, Open in new tab and Remove link buttons.
• Drag handle. Hover over a block to show its handle. Drag the handle to move the block, or select it to choose Turn into (another block type) or Delete.
• Table tools. With the cursor inside a table, a small toolbar lets you add a column before or after, delete a column, add a row above or below, delete a row, or delete the whole table.
• Pasting Markdown. Markdown you paste is converted into formatted blocks.

Draft with the Guidelines Assistant

Import a template

Arabic Agents and Voices

The Arabic Language Lock

Arabic Accent

• Picking a regional option pins the transcription engine to that variety — noticeably fewer misheard words with strongly regional callers.
• Arabic — any region and Fusha transcribe exactly as Arabic always did, so choosing either changes nothing about recognition.
• On the speaking side the provider accepts a narrower set of hints, so Emirati and Fusha change delivery directly while the other regional choices take effect through better recognition.

Voice tuning for Arabic voices

Knowledge Base

Actions

Pre-conversation Actions

• Runs before the agent speaks.
  
• Common use-case: FetchContact – pull CRM details using an API Request node and store them in variables like {{firstName}}, {{accountStatus}}.
  
• Variables created here are available to personalize the Greeting message and Guidelines.
• Variables values are available throughout the lifetime of the interaction and can be utilized in In-conversation or Post-conversation Actions.

In-conversation Actions

• Triggered on demand when the caller asks for something (e.g. “CheckOrderStatus”).
  
• Each execution is logged in the live transcript and action logs, giving you full visibility into what happened during the conversation.
  
• Provide comprehensive documentation for each action, including specific trigger conditions and use cases. Detailed context helps the AI model accurately determine when and how to execute the action.
• Define parameters to specify the data requirements for action execution. Each parameter can be configured with:
  • Parameter Name
  • Description (use this to provide format requirements and context to help the AI collect data correctly)
  • Type (String, Number or Enum – with Enum, list the allowed values under Enum Items using Add Item)
  • Required switch (Yes or No) Parameters are optional and you can configure zero or more as needed for the action.
• Once an action has a parameter, Assign Parameter Values to Variables lists each one. Pick or create a variable to hold what the AI collects, and optionally enter a default value to use when the AI doesn’t supply one.
• You can define filler messages that the agent speaks while waiting for the API response using the Say node in the flow builder
• External API calls can be executed through the flow’s API Request Node

Post-conversation Actions

• Generating a conversation summary and disposition
• Saving the transcript to your CRM or data warehouse
• Kicking off downstream workflows (e.g. ticket creation using flow nodes)

Agent Settings

Languages & localization

• Multilingual – English stays the primary language. Use Add another language to add any of Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch; they appear under Additional Languages, and the × beside one removes it. With none added, the panel reads English is currently the only language enabled.
• Monolingual – the agent works in a single language, chosen under Primary Language, and English responses are turned off. English isn’t in the Primary Language list, so for an English-only agent use Multilingual with no additional languages.

• When you add a language, Studio translates the agent’s existing messages into it in the background and shows New languages added! Please check back in a few minutes for translated messages to be available.
• Beside each language’s box is a wand button whose tooltip reads Translate to followed by the language. It translates the English message again on demand, and needs an English message to translate from.

Summary Language

• Pick any of the agent’s configured languages, or Auto (detect from conversation) to match the language the caller actually used. Defaults to English.
• The whole summary is written in that language — including the Additional Concerns section.
• The four section headings (Issue, Resolution, Next Steps, Additional Concerns) are written in that language too, in the same order and bold layout. In English they read as shown here. In another language each heading is translated — in Spanish, for example, Problema:, Resolución:, Próximos pasos: and Comentarios adicionales: — and the wording of a translated heading can vary slightly from one summary to the next.

Timezone

Transcription Keywords

• You can add up to 100 keywords per agent, and a running Total: n/100 counter shows how many are in use. A keyword that is already on the list (ignoring capital letters) is rejected.
• A call can use about 500 words in total across all of your keywords. If the whole list is longer than that, none of the keywords are applied to the call, so keep each keyword short.
• Keywords aren’t available while the agent is Multilingual with at least one additional language. The card is greyed out and reads Keyword-based transcription isn’t available for multilingual AI agents. They work for Monolingual agents, and for Multilingual agents that have no additional language.

Max Call Duration

• Max Duration Message – what the caller hears when the limit is reached. A new agent’s message reads “Exceeded the max call time.” Each additional language has its own version, translated for you when you add the language. It is a plain text box, so write the message exactly as you want it spoken.
• Any Post-conversation Actions still run after the call ends.

Toggle Transfer

Silence Detection

• Silence Timeout – how long the agent waits before saying something, from 10 to 45 seconds in 5-second steps. A new agent starts at 10 seconds.
• Max Retries – how many times the agent checks in before it hangs up, from 1 to 5. A new agent starts at 3.
• Silence Detection Message – what the agent says to prompt the caller. A new agent’s message reads “Are you still there?” The agent speaks it exactly as you write it, so it is the words themselves, not an instruction to the model. Each additional language has its own version, translated for you when you add the language.

• The timer counts the caller’s silence. It waits while the agent is thinking, speaking or running an action, and any caller speech or keypad press resets it.
• Each time the timeout passes, the agent speaks the Silence Detection Message. With the default Max Retries of 3, the agent checks in after the first, second and third timeouts. After the next timeout it plays the Closing message from Greeting & Sign-off and ends the call.
• If the caller speaks in between, the count of check-ins starts again.
• Call History records a call that ended this way with the disposition Hangup.

Noise Cancellation

• What it costs. It’s a per-minute add-on, charged on top of your voice rate for every call the agent handles. When the card can show your rate, it does. Otherwise it says only that you will be charged per minute.
• Turning it on. The switch opens Turn on noise cancellation?, which repeats the charge. Select Turn on to confirm, or Cancel to leave it off. Turn on waits until the rate has loaded. Turning it off needs no confirmation.
• When it applies. Either change saves on its own and applies from the agent’s next call. It doesn’t wait for a Publish, even on a published agent.
• Balance. A call is refused when the balance can’t cover it.
• What is charged. Only the agent’s own part of a call. The time after a transfer to a person isn’t charged, and neither are Talk test calls.
• Finding it. Press Cmd/Ctrl+K, type “noise” and choose Noise cancellation under the agent’s heading.
• The Co-pilot. Ask the Co-pilot to turn it on and it first shows a card that says it adds a per-minute charge to every call and carries This spends money. Select Yes, go ahead to confirm. Turning it off needs no card.

Greeting & Sign-off and Handoff

Greeting & Sign-off

Handoff

Outcome options

Transfer settings

• Transfer Message – what callers hear while the call is transferred. It is labelled Message to Play in English (Default), with another box for each additional language, labelled Message to Play in and the language. If the caller hasn’t spoken yet when the agent decides to transfer, for example because they have only pressed keys, a Multilingual agent plays the message in English: your Message to Play in English (Default), or a built-in line if that box is empty. Once the caller has spoken, it plays in the language they used. Type {{ to insert a variable. This message can use interaction.DISPOSITION and interaction.TRANSFER_DESTINATION as well as the usual call variables. See Built-in variables.
• Transfer Rule – where the call goes. Choose a Destination Type, Phone Number (the default) or SIP URI, then enter the destination: a fixed value, a variable, or {{interaction.TRANSFER_DESTINATION}} to let the agent work it out from the conversation. SIP Integration describes every setting, including the Screen Pop and Inline SIP context headers switches that SIP URI destinations offer. If Transfer Rule isn’t on the Voice tab, contact your provider.

Phone Numbers on an Agent

Variables Manager

• Find a variable with Search variables.
• Create, edit, and delete variables with descriptions. Each row has View references, Edit variable and Delete variable buttons.
• Filter for unused variables with the Unused button to keep the agent tidy.
• Open View references on any variable to see exactly where it is used — across pre-conversation, in-conversation, and post-conversation actions (down to the flow node) and assistant settings.
• Variables that are still referenced cannot be deleted, so flows never break silently. Studio says Delete is blocked while this variable is still used by the AI Agent. and lists the references to remove first.
• A variable that a published agent’s live chat widget still maps can’t be deleted either, even after you remove the mapping in Chat Configuration, until you publish the agent. Studio shows Cannot delete variable while it is still used: the published chat widget of ”…” maps it. Publish the agent first, then delete the variable.
• On a published agent, creating, editing or deleting a variable makes the agent read Unpublished edits, and a new name, description or default value reaches live conversations at your next Publish.

Creating a variable

• uses only letters, numbers and underscores;
• starts with a letter or an underscore, and doesn’t end with an underscore;
• is up to 100 characters (keep a Chat widget variable’s name to 64 or fewer if you want to map it in the widget’s settings);
• isn’t email (in any capitalization) for a Chat widget variable, because the platform already provides the visitor’s email address as {{interaction.CHAT.email}} and it can’t be created — see Skip the Email Screen;
• is different from every other variable on the same agent. Studio refuses a repeated name and says that it already exists.

Built-in variables

Default values

Which value wins

Behaviour worth knowing

• A default can be up to 2,000 characters. A blank or whitespace-only entry is stored as no default at all.
• A variable has no default until you fill the field in.
• The edit drawer always submits the field, so saving an edit with Default value empty clears a default you set earlier. Re-enter the value if you only meant to change something else on the row.
• A variable created on the fly from the JSON Path Expressions table of an API Request node inherits that row’s Default Value (optional) as its own default.
• A Decision node that tested a variable for emptiness may take a different branch once that variable has a default, because the variable is no longer empty at the start of the interaction. Review those conditions when you add a default to a variable a flow branches on.
• On a published agent, a new or changed default reaches live conversations at your next Publish. Talk and Chat use it at once.

Autosave & Draft Safety

• Changes to the Greeting message, Closing message, Guidelines, voice selection and tuning, Handoff outcomes and transfer settings, and Agent Settings save automatically within a second or two after you pause typing.
• An agent’s name is the exception. Rename it with the pencil icon in the agent’s header and confirm with Enter or the check mark — see The Agent Header.
• The panels for Greeting & Sign-off, Voice Configuration, Handoff and Agent Settings have no save button. A voice you pick saves at once, and every other change in those panels saves a moment after you stop editing.
• A toast in the bottom-right (for example “Voice updated”, “Voice settings updated”, “Settings updated” or “Saved changes” for Guidelines) confirms each successful save to the draft. If something fails you will see an error toast so you can retry. The chat greeting’s confirmation (“Chat greeting saved”) appears only after you edit it; opening an agent doesn’t show it.
• If the Co-pilot changes the agent while you have an edit waiting to save on Agent Settings or Handoff, both changes are kept. If you both changed the same setting, yours is saved. The agent’s languages count as one setting: if you changed its languages or language mode, your whole language setup is kept.
• An open Guidelines editor or Voice Configuration panel switches to the Co-pilot’s version as soon as it lands. The Co-pilot only sees what has been saved, so give a hand edit a moment to save before you ask it to build on that edit. Its own edits go to the draft too, so on a published agent they wait for your next Publish, and it can’t publish for you.

Publishing Your Agent

• Before the first Publish, callers, chat visitors and email customers get your latest saved edits, so a change is live as soon as it saves. A new agent works this way, and so does a duplicate or an imported agent, which starts as Not published.
• After the first Publish, phone calls (including a campaign’s Test Call and on-demand calls), chats on your website and emails run the published version, and later edits wait for your next Publish. A call that is already in progress isn’t affected.

Publish Step by Step

What Waits for Publish

• The greeting and closing messages, Guidelines, voice and voice tuning, and the agent’s name.
• Agent Settings, other than noise cancellation.
• Handoff: the transfer message and destination, its switches and the outcome options.
• Actions and their flows, variables with their descriptions and default values, and data transformers.
• Linked knowledge groups, FAQs and changes to a group’s content.
• Everything on Chat Configuration (the widget’s name, logo, avatar, layout, colors, disclaimer, proactive notification and chat variables) and on Email Configuration (Inactivity Timeout and AI Disclaimer).

• The phone numbers assigned to an agent.
• The mailboxes it’s ticked for on Email Configuration → Inbox.
• The chat widget’s Allowed Origins and Widget Key.
• Noise cancellation.
• What an authentication profile holds, such as its credentials. Which profile an action uses still waits for Publish.

• Choosing a different Handoff Platform points escalations at the new connection straight away, but brings in its channel list only at the next Publish, so publish right after changing it. See Chat Escalation.
• A variable that your live chat widget still maps can’t be deleted until you publish without the mapping. See Variables Manager.
• The Co-pilot works the same way. Its edits go to the draft, it can’t publish, and its confirmation cards for the transfer destination and Handoff Platform say when a change waits for Publish. Press Publish yourself afterwards.

When a Publish Stalls

Version History

• Draft – your working copy. Its badge shows the number the next Publish would give it, and it reads Unpublished changes or No changes.
• The live version, marked Live.
• Older Versions, newest first.

• View Changes – compares this version with one other. The version you opened it from stays fixed, and you pick the other side from the dropdown after with: your draft, the live version or any other version. It starts on the live version (on the live version’s own row, the version before it) and shows what changes going from the side you picked to the version you opened it from, in the same sections as Review Changes. The Draft row has only View Changes, which is greyed out until the agent has a live version.
• Make this version active – sends calls, chats and emails to this version. See Make an Earlier Version Live. The live version doesn’t offer it.
• Export Version – downloads that version as an agent file, as it was when you published it, and confirms with Export downloaded. The file’s name ends with the version number, such as -v3.json. Select Import Agent on Build → AI Agents to bring it back as a new agent that starts as Not published. See Export, Import and, for how its knowledge groups come back, Import and Export.
• Delete Version – after you confirm, removes the version for good. The knowledge groups and chat widget images it used are released, and any that no other version or your draft still uses are deleted. Calls are not affected. The live version and an agent’s only version can’t be deleted. See Knowledge Groups and Published Agents.

Make an Earlier Version Live

• What callers will see – what changes going from the live version to this one.
• Your next Publish – what your next Publish would review once this version is live: this version against your draft.

• Anything listed under This version cannot be made active yet has to be fixed first, for example when the voice or a knowledge group the version uses is no longer available (or the group’s files are still being added), or an authentication profile it uses no longer exists. Fix the cause, or pick another version.
• Each warning under Tick each one to continue has to be ticked. Warnings cover things such as a chat widget image that is no longer stored, integration actions that are turned off or have no connected account, a human handoff connection that is no longer connected, and campaigns that are live or paused, whose next calls will run this version.

• New phone calls, chats on your website and emails run that version, and so do the next calls of a running campaign. A call that is already in progress isn’t affected, and a chat or email thread that is already open may pick up the new version when it next continues.
• Your draft isn’t changed. Talk and the test Chat keep using the draft.
• The version’s row gains a Made live by line that says who did it and when.
• The agent’s header usually reads Unpublished edits, because your draft still differs from the version that is live. Your next Publish reviews the draft against that version and puts the draft live as a new version.
• Settings that apply at once and aren’t part of a version keep their current values. See What Waits for Publish.

Testing Your Agent

Noise Cancellation on a Talk Call

• The status line reads Switching noise cancellation… and then Noise cancellation on.
• Select the button to turn noise cancellation off, and the line reads Noise cancellation off. Select it again to turn it back on.
• It changes this test call only, never the agent’s own Toggle noise cancellation setting, and Talk calls aren’t charged.
• A Talk call starts with noise cancellation on, whatever the agent’s own setting. Phone calls follow the agent’s setting and can’t be switched mid-call.
• If noise cancellation can’t start or drops, the call carries on with the original audio and the line reads Noise cancellation isn’t running, so this call is using the original audio. Select the button to try again. If a switch isn’t confirmed within about 8 seconds, a message says Noise cancellation didn’t switch on. Please try again. (or off).
• After a transfer, the line reads Noise cancellation ended when the call was transferred and the button goes.
• The status line and the button come with the call’s live transcript, so they don’t appear if the transcript can’t connect.

The Test Chat

• Press Enter to send a message, or Shift+Enter to start a new line.
• A message can be up to about 8,000 bytes. That is roughly 8,000 characters in English, and fewer in languages whose characters take more than one byte. A longer message isn’t sent. Studio shows Message is too long. Please shorten it. and keeps the send button off until you shorten it. The live chat widget has the same limit.
• Close the test chat and reopen it within 10 minutes of its last activity, and you carry on with the same conversation. After longer than that, you start a new one. New session starts a new conversation at any time. It also appears beside Session ended. Start a new session to continue. when a conversation ends.

Caller ID on Transferred Calls

• If the caller’s number is withheld, or the call didn’t come in from a phone number (a browser call or a SIP call, for example), the destination sees the agent’s own number instead: the earliest-assigned of its numbers bought in the Studio. The same rules apply as in Testing Transfers below: BYO numbers and numbers with an unresolved payment failure are never used.
• If the agent has no qualifying number, the transfer may fail. Assign a number bought in the Studio to the agent before you rely on transfers from those calls.
• Transfers to a SIP address that use SIP REFER are unchanged.

Testing Transfers

• Numbers you brought yourself (BYO) are never used.
• Numbers with an unresolved payment failure are never used either. When a subscription payment fails, every number you bought is flagged, so none of them qualifies until the payment is resolved. See When a Payment Fails.
• The Assigned numbers chip in the agent’s header lists the most recently assigned number first, so the caller ID may not be the first number it shows.
• A transfer to a SIP URI from a Talk call also carries the Screen Pop token and the Inline SIP context headers when you have turned them on in Handoff. You can try the receiving system end to end without a real phone call. See Screen Pop. A Talk call has no real caller number, so the caller details on the Screen Pop page may look odd.

Copying, Duplicating & Transferring Agents & Actions

Duplicate & Copy Actions

• Duplicate adds a copy to the same agent.
• Copy to Agent opens a drawer where you pick another agent in the same account and select Copy. Studio then opens that agent’s list of the same kind of actions. The drawer notes: “Referenced variables are created in the target when missing. Campaign variables are not carried over and are removed from the copy.”

• Duplicate names the copy Copy of followed by the original name. An in-conversation action’s name can’t contain spaces, so check_order_status becomes Copy_of_check_order_status.
• Copy to Agent keeps the original name when the other agent has no action with it.
• When the name is taken, the copy gets the next number: _2, _3 and so on for an in-conversation action (check_order_status_2), or (2), (3) for a pre- or post-conversation action (FetchContact (2)).

Duplicate

• Open the More actions menu (...) on an agent’s row and choose Duplicate to create a Copy of {Agent Name} with the same actions, flows, variables, outcomes, linked knowledge groups, and data transformers. If that name is taken, a number is added, such as Copy of Support (2). Studio confirms with AI Agent duplicated successfully. The copy is made from your latest saved edits, including any you haven’t published, and it starts as Not published, so it runs its draft until you publish it.
• The copy gets its own actions, flows and variables, so you can rename it and adjust those without touching the original agent.
• Knowledge Groups and Data Transformers are shared with the original, not copied. The copy links to the same Knowledge Groups and uses the same Data Transformers, so editing a transformer, or a group’s content, changes it for both agents, though a published agent’s callers get the change only at its next Publish.

Export

• From the same menu select Export to download a JSON bundle that includes the agent, actions, flow definitions, knowledge links, dispositions, data transformers, and usage mappings. Studio confirms with Export downloaded. The bundle holds your latest saved edits, including any you haven’t published. To export a version as it was when you published it, use Export Version in Version History.
• The file is named assistant- followed by the agent’s name and .json. The name has no date or time in it, so exporting the same agent again produces the same file name — rename earlier backups if you want to keep versions. The time of the export is recorded inside the file.

Import

• Select Import Agent, beside Create Agent at the top of Build → AI Agents, and upload a bundle that was exported from SquawkVoice Studio. The file must be a .json file of up to about 5 MB. The button reads Importing Agent… while it works.
• The importer recreates flows, restores data transformers, and reconnects or rebuilds knowledge groups. Studio shows AI Agent imported successfully when it finishes, or Failed to import AI Agent if it could not. A file that is too large, or isn’t an agent export, fails this way.
• The new agent is named Import of {Agent Name}. If that name is taken, a number is added, such as Import of Support (2). It starts as Not published.
• Knowledge Groups depend on where you import. Into the workspace the agent was exported from, the agent keeps its links to the same Knowledge Groups. Into a different workspace, each group is matched by its content and then by its name. A group with no match is rebuilt from the bundle: its URLs and FAQs are recreated, and its uploaded files too when you import within the same region. A file from Export Version follows one rule wherever you import it: a group is reused only if it is still in the workspace with the same FAQs, and otherwise a new group is created, never matched by name. See Import and Export for what to check afterwards.
• Data Transformers follow the same split. Into the workspace the agent was exported from, the imported agent uses the same transformers as the original, so editing one changes both, and any transformer that has since been deleted is created again. Into a different workspace, each transformer is compared with the ones already there. One whose Expression Pipeline is exactly the same is reused, even under a different name. Any other is created as a new transformer, with a number added to its name if the name is taken, such as Format Phone_(2). See Data Transformers.

Best Practices

• Start simple. Launch with a lightweight set of actions and expand iteratively.
• Test, then publish. Try a change with Talk and Chat, check it in Review Changes, then Publish. See Publishing Your Agent.
• Keep guidelines concise. The shorter the system prompt, the faster the agent responds.
• Monitor analytics. Use the built-in dashboard to track resolution rates, average call duration, and customer sentiment.
• Iterate on feedback. Real-world conversations reveal edge cases you can cover with new actions or guideline tweaks.

Next Steps