Release notes

An Agent can now hold up to 20 active change sets at a time. The previous limit was 5.

The 10-entities-per-change-set limit is unchanged. Creating a twenty-first active change set still returns an error; promote, revert, or discard one to free a slot.

Playbooks now support flow controls that let you change how the AI Agent responds when an end user deviates from the current flow.

New feature

Three controls are available under Advanced settings in the Playbook details pane:

• Asks for human: Choose whether the Agent exits the Playbook or stays in the flow.
• Goes off-topic: Choose whether the Agent suspends the Playbook to address the new topic (then returns or exits) or stays in the flow.
• Asks a Q&A question: Choose whether the Agent searches Knowledge or stays in the flow.

A Playbooks SET step in Use reasoning mode can now resolve up to 5 variables. The AI Agent reads the transcript once and extracts every value in a single pass. The step stays silent, so the end user is never prompted.

The step keeps one instruction, which applies to every variable. Each variable is configured as an input, and keeps its own accepted values and fallback value. A variable the Agent cannot resolve takes its fallback value. When no fallback value is set, that variable keeps the value it already held.

Code tools let an AI Agent run a small piece of sandboxed Python during a conversation to reshape data, compute a value, or return only what it needs. A code tool does everything an API tool does — call an external endpoint and get data back — and can also run logic on the result before the Agent uses it.

Code tools are authored via Ada’s MCP Server or the Platform API. Over Ada’s MCP Server, you describe the tool to an AI assistant and authoring goes through a change set (describe, generate, stage, test, then promote), so a tool never affects live conversations until it is approved. A push through the Platform API writes the tool directly, with no change-set step. Add a tool to a Playbook step for the Agent to run mid-conversation (recommended), or invoke it directly like an API tool.

A Playbooks ASK step that maps the answer using reasoning can now set up to 5 variables from a single end user response. The AI Agent asks for the values together and extracts all of them in one pass. If the response provides only some of the values, the Agent saves those and re-asks for the ones still missing. The end user can correct any captured value before the step completes.

The Tools API is a single inventory of the HTTP tools (api) and code tools (code) configured on an AI Agent. List, fetch, create, update, and delete tools through one Platform API.

API endpoints

The following REST endpoints are available:

• List tools (GET /v2/tools/): Returns the Agent’s tools in cursor-paginated pages, in source order api, then code. Filter with type and enabled. Replay meta.next_page_url unchanged to fetch the next page.
• Get a tool (GET /v2/tools/{tool_id}): Returns a single tool by its id, including the type-specific body (api or code). Authorization headers are never returned.
• Create a tool (POST /v2/tools/): Creates an api or code tool.
• Update a tool (PATCH /v2/tools/{tool_id}): Partially updates a tool. api and code tools accept contract, control, and implementation writes.
• Delete a tool (DELETE /v2/tools/{tool_id}): Soft-deletes an api or code tool.

The Topics & Intents report’s CSV export now includes four new columns after the name columns: Topic Description, Intent Description, What’s Included, and What’s Excluded. This gives reviewers and external stakeholders the full definition of each Topic and Intent without opening them in the dashboard.

Because the new columns are inserted before the metric columns, spreadsheets or scripts that read the export by column position need to be updated; reading by header name is unaffected.

Ada removed the scripted proactive campaigns feature and deprecated the campaign actions in the Embed2 Web SDK. Existing integrations continue to load without code changes, but the deprecated actions no longer have any effect.

What changed

• The triggerCampaign, evaluateCampaignConditions, and trackEvent actions are deprecated. Calling them still resolves, but has no effect and logs a console warning. To start a proactive conversation, use triggerProactive with a generative AI Agent, or triggerAnswer with a scripted bot.
• The ada:campaigns:shown and ada:campaigns:engaged events are unchanged. The ada:campaigns:opened event no longer fires. See subscribeEvent.
• Automatic URL-triggered business events are unchanged: Embed2 still tracks configured business events when a page URL matches their trigger conditions.

Your Voice AI Agent can now change how fast it speaks when a caller asks it to.

What’s new

A caller who finds the AI Agent too fast or too slow can say so, and the AI Agent adjusts its delivery and continues the conversation.

• Triggered by the caller. Phrases like “you’re talking too fast” or “could you speak a bit faster?” move the pace one setting along Slowest, Slow, Normal, and Fast. Callers can ask again to keep moving in the same direction.
• Applies to that call only. The Pace selected for each language is unchanged, and the next call starts from it again.
• Changes delivery only. The AI Agent says the same thing and stays on the same step of its Playbook.

Your Voice AI Agent now speaks more slowly when it reads a captured value back to a caller to confirm it.

What’s new

Read-backs are the moment a caller is most likely to ask “sorry, can you repeat that?” — a slower delivery makes order numbers, email addresses, and dates easier to verify on the first pass.

A Playbooks ASK step can now build its accepted values from a variable holding a list instead of a hand-typed list, so the options match the individual end user’s data — for example, one option per open order. Turn it on per variable with Pull from data source in the step’s accepted values, then choose the variable that holds the list.

The list is resolved each time the step asks. If the data source is empty or unreadable, the step falls back to accepting free text rather than failing.

Behavior changes made through the MCP server no longer apply to live conversations as soon as they are saved. They are staged on a change set, can be tested against simulations first, and go live only after explicit approval. A change set that has gone live can be reverted in one step.

What’s new

Change sets add a draft, test, promote, and revert loop to Agent behavior changes.

Testing an AI Agent over voice previously meant having it call your phone. You can now run the call in your browser instead, and voice testing is available to every AI Agent with Voice enabled.

Choose how to connect

The Test AI Agent button now offers Chat or Voice. Selecting Voice opens a setup bar where you choose how to be reached before anything connects:

Added a Trusted senders list to the Email channel. Email from a trusted sender reaches the AI Agent’s normal flow — AI-powered auto-reply detection no longer dismisses it — so system mail like helpdesk auto-acknowledgments and quarantine digests can be triaged by your playbooks and handed off to a human.

Manage the list from the new Trusted senders section on the Config > CHANNELS > Email page. Header-based auto-reply detection, content moderation, and loop protection still apply to trusted senders, and an address can’t be on both the trusted and blocked lists.

In Messaging and Email channels, when a Knowledge search finds no articles matching an end user’s localized language tag, your AI Agent now uses the remaining articles in that language, including versions localized for other regions. Previously the search returned nothing and the AI Agent replied that it did not have the information. Voice is unchanged.

A Playbooks ASK step that maps the answer using reasoning can now present its accepted values as tappable quick replies, so end users select an option instead of typing one. Turn it on per step with the Quick replies option in the step’s Messaging channel settings. It is available when the step sets a single variable and has at least two accepted values.

You can now control how fast your Voice AI Agent speaks in each language.

What’s new

Each language on the Config > CHANNELS > Voice > Speaking voice tab now includes a Pace slider with four settings: Slowest, Slow, Normal, and Fast.

• Nothing changes for existing Agents. Every language defaults to Normal, which matches the previous speaking speed.
• Set the pace per language — for example, slow down one language without changing the others.
• Preview before saving. Type a text sample and click Play to hear the selected voice at the selected pace.

Channels in the Conversations API now carry a capabilities field that declares what the channel’s surface can render, and the AI Agent tailors its replies to match. markdown is the first capability: a channel that opts in receives replies written with bold, italics, and inline links instead of plain prose.

Capabilities are disabled by default, so existing channels are unaffected until they opt in. Set capabilities when creating a channel with POST /v2/channels, or change it on an existing channel with the new PATCH /v2/channels/{channel_id} endpoint. Native channels expose capabilities as read-only, derived from their surface.

Playbooks can now be assigned to conversation start events so they run automatically when a new conversation begins on a specific channel. Supported channels include Messaging, Voice, Email, and all configured Sunshine social channels.

In the Playbook editor’s Details pane, type @ in the When to use this Playbook field to assign one or more events. Each event can only be assigned to one Playbook at a time.

Images delivered by an AI Agent can now be chosen at runtime instead of at authoring time. A single Playbook can send a different image per conversation — an order-specific product photo, for example — rather than requiring near-duplicate Playbooks.

Media Variables

A new Media data type is available in the Variable Manager. It stores a publicly reachable http or https image URL.