Release notes

This release adds commands that draft and create the objective of a changeset. An objective is the metric that measures one changeset. The commands show only on an AI agent that has the improvement loop turned on.

Changes

This section lists what the release adds or changes.

Use reasoning SET steps in Playbooks can now store an answer from Knowledge in a variable. The step sends no message to the end user. A later step can check the variable and decide what the AI Agent sends or passes to another system.

What’s new

Two options appear under Knowledge in the @ menu of a Use reasoning SET step that resolves one variable.

You can now connect a SharePoint Online folder to Knowledge. Each connection imports the Word, PDF, PowerPoint, Excel, CSV, HTML, Markdown and text files in one folder and its subfolders into its own source, and syncs changes every 15 minutes. Your Microsoft admin registers a Microsoft Entra app with read access to the site, and the folder doesn’t need to be public.

An owner or an admin can now set up the outbound SIP trunk for Voice Proactive Outreach in the dashboard. The trunk carries the calls that your system starts with the Conversations API. Until now, Ada configured the trunk with you.

The new Outbound calling section in Config > CHANNELS > Voice > Configuration holds the trunk settings: the termination host, the transport, the authentication, the caller ID and the edge. After you save, place a test call to a phone number that you choose and read the result in the section. Then turn the trunk on. A test call is a real call, and your carrier bills it.

You can now give a web import a list of CSS selectors. Before each page becomes an article, the scraper removes the elements that match them, such as a site footer that repeats on every page. The selectors apply from the next sync of the source.

A source takes up to 20 selectors, of up to 75 characters each. You cannot change them while a sync of the source runs. If the selectors would remove all of the content of a page, the page keeps its last imported version.

This release lets you set the reasoning level of staged playbooks, and warns you when a staged playbook would lower a step from High reasoning to Normal.

Changes

This section lists what the release adds or changes.

• changesets stage takes --reasoning-level <high|normal>. It sets the reasoning level on every set and ask step in the staged playbooks, and overrides the levels in the file. For example: adacli <instance> changesets stage --name "Orders on High" --file changes.json --reasoning-level high.
• changesets stage warns when a staged playbook would change a step from High reasoning to Normal. The warning names each step. The change still stages.

This release warns before testing runs create starts a batch that may not fit in the daily test run quota.

Changes

This section lists what the release adds or changes.

• testing runs create reads the daily test run quota before the first run starts. A batch is the number of unique test case IDs times --repeat. If the batch is larger than the runs left today, the command prints a warning on stderr: the batch may not fit, and some runs can start before the api refuses the rest. The command then starts the runs, and the api decides. The exit code and the --json output do not change.
• testing runs create --no-quota-check starts the runs without a read of the quota. For example: adacli <instance> testing runs create --cases <id> --repeat 3 --no-quota-check.
• testing quota and testing runs create print a warning when the daily limit is far below the default of 3,000 runs. Contact Ada support if you see it.

This release makes playbooks create create an inactive playbook unless you pass --active.

Changes

This section lists what the release adds or changes.

• playbooks create makes an inactive playbook by default. An inactive playbook does not run in conversations with end users. Before this release, a playbook created with no state option was active at once. If a script relies on that, add --active to the command.
• playbooks create --inactive states the default. For example: adacli <instance> playbooks create --name "Returns" --file steps.json --inactive. --active and --inactive cannot be combined.
• When the new playbook is inactive, the command prints how to make it live: playbooks update <id> --active.

This release adds commands that test a changeset against the live AI agent. They show only on an AI agent that has the improvement loop turned on.

Changes

This section lists what the release adds or changes.

• changesets test start <changeset_id> --metric <id> runs each test case that lists the metric on the changeset and on the live AI agent. The metric is the objective of the changeset or a custom metric. --repeats sets the runs of each test case on each side, from 1 to 5. The default is 3. The runs use Ada’s improvement quota, not the simulation quota. For example: adacli <instance> changesets test start <changeset_id> --metric <id>.
• changesets test get <test_id> shows the result: the outcome of each test case on each side, and the test cases that the changeset fixed and broke. For example: adacli <instance> changesets test get <test_id>.

You can now connect Microsoft Copilot to the MCP server yourself. Build a Copilot agent in Microsoft Copilot Studio, and add the MCP server with your OAuth connection URL. People then use the Copilot agent in Microsoft 365 Copilot and Microsoft Teams.

Previously, Ada packaged a Copilot agent for each AI Agent, and an IT admin imported the package into Copilot Studio. You no longer need the package. Copilot agents imported from a package continue to work. An admin is necessary only to make the Copilot agent available to the whole organization.

The iOS, Android, and React Native Messaging SDKs now accept branding to override dashboard appearance for each SDK instance. An app can set its own AI Agent name, avatar, colors, and text size with the same fields as the web SDK. Fields that you omit keep their dashboard values. Overrides stay in effect after a chat reset and document reload. Branding does not add to the WebView URL length. Malformed input is ignored and reported once per document without its value. Host overrides, including launcher text and icon URLs, are not saved in the startup cache.

You can now leave pages out of a web import that crawls an entire website. Under Every webpage starting from one URL, add the paths to skip in Exclude pages, separated by commas. A path such as /blog excludes that page and every page under it, and * matches any text.

Set the list when you add a website, or change it later from the source’s Settings on the Sources tab. A change applies from the next sync. Articles already imported from a page you exclude stay in your knowledge until you delete them or set them to Inactive.

This release lets you set the domains that a code tool can reach.

Changes

This section lists what the release adds or changes.

• code-tools allowed-domains manages the domains that a code tool’s fetch() can reach. A fetch() to a domain that is not on the list fails.
  • add <domain> adds a domain. A domain is a scheme and a host. It must start with https:// or http://. It can have a port, and it can start with a *. wildcard for one subdomain. For example: adacli <instance> code-tools allowed-domains add https://api.example.com.
  • list shows the ID and the domain of each entry.
  • remove <id> removes a domain. Get the ID from list.

A code tool can now send an email with send_email(to, subject, body, bcc=None, html=False, reply_to=None). Use it to put a one-time code, an order confirmation or a receipt in the end user’s inbox during a conversation. The result tells the Playbook whether the email was sent, so the Playbook can branch on it.

What’s new

Each save of a code tool now creates a numbered version. You can compare two versions and restore an earlier one from the code tool’s page.

What’s new

Open a code tool, open its menu, and select Version history.

• Every save is recorded. A version is created when the tool changes in the dashboard, over the Platform API, or when a change set promotes it. A test run does not create a version. Ada keeps the newest 25 versions of each tool.
• Compare two versions. Select a Version and a version to Compare with. The Changes tab shows each changed field, including the code, inputs, outputs and Availability rules. The Full version tab shows the whole tool as it was, read-only.
• Restore a version. Select Restore version N. Before anything changes, a confirmation shows what the restore changes in the current version. The restore brings back the version’s code, inputs, outputs, Availability rules and Block in simulations setting. It keeps the current environment and Active setting. The restore itself is saved as a new version, so you can undo it.

Starting November 10, 2026, Ada is gradually updating how it assesses automated resolution to improve accuracy. Ada starts using the updated assessment for each AI Agent at a different time, and the AI Agent’s automated resolution rate may go up or down from that day. Conversations from before that day keep their original classification.

What isn’t changing

Lovelace 1.21.4 changes the DropZone component of the @ada-cx/lovelace design system.

• DropZone no longer renders a hidden keyboard drop button. That button was a Tab stop, and focus on it opened the drop panel.
• The drop panel shows only while a drag is over the zone.
• label is now the panel heading only.

If your interface uses DropZone, also give a control that opens a file picker. Keyboard and screen reader users attach files with that control. See DropZone.

Lovelace 1.22.0 deprecates three component tokens of the @ada-cx/lovelace design system. Lovelace components no longer read these tokens, so an override of them no longer changes message text.

• message-content-default: use color-content-default. For the text in a user message, use message-user-content.
• message-content-font-size-body: use font-size-body.
• message-content-line-height-body: use line-height-body.

This release lets you set the AI Agent settings from the CLI: persona manners, the agent name, the avatar and image redactions. You must have the Admin or Owner role to change them.

Changes

This section lists what the release adds or changes.

• client persona update takes an on/off flag for each manner: --emoji, --ask-for-more-help, --handoff-friction (offer a human agent only after trying other options), --avoid-customer-pi, --avoid-competitors and --acknowledgment-messages. Add --no- to turn one off, for example --no-emoji.
• client persona update also takes --allowed-emojis <emojis>, --competitor-names <csv>, --contact-details <none|all|approved_list> and --contact-detail-values <csv>. For example: adacli <instance> client persona update --avoid-competitors --competitor-names "Acme,Globex".
• client persona show shows each of these settings. It no longer shows an Instructions line, which was always empty.
• client update sets the AI Agent name and the avatar. It takes --agent-name <name> with an optional --language <code>, and either --avatar <file> or --avatar-letter <letter>. For example: adacli <instance> client update --agent-name "Max" --avatar ./avatar.png.
• redactions images sets which documents in an image the AI Agent does not read. list shows each class. enable <class> and disable <class> change government_id or other_credential. Payment card images are always withheld. message set <text> and message clear change the message the AI Agent sends when it withholds an image.

A new Send progress messages during long tool runs setting lets the AI Agent keep the end user informed while a Playbook RUN step waits on a slow API tool or Code tool. With the setting on, the Agent sends a short message after the tool has run for 10 seconds. It sends another at 20 seconds. Then it sends one every 20 seconds until the result arrives. Each message says the Agent is still working, in the language of the conversation. A tool that returns within 10 seconds sends nothing.