Release notes

The Simulations feature (formerly Testing at Scale) now runs test cases as multi-turn conversations by default. This brings dashboard testing closer to real end-user interactions and expands what can be validated before deployment.

What’s new

• Multi-turn by default: Every test case runs as a multi-turn simulation, up to 40 turns per simulation.
• Scenario field: Each test case now requires a Scenario that guides the simulated end user’s goal, context, and responses across turns.
• Up to 10 expected outcomes: The per-test-case limit on expected outcomes increased from 3 to 10.
• Multi-turn execution via MCP Server: Simulations executed through the MCP Server now run as multi-turn conversations. Creating and updating test cases (including Scenarios and Expected outcomes) via MCP is also supported.

Greetings are still excluded from simulated conversations. Other limitations are unchanged.

Existing test cases remain runnable. The Customer inquiry set on each test case is used as the Scenario until the Scenario is updated. The next time a pre-existing test case is edited, updating the Scenario is required before saving.

Learn more in the Simulations documentation.

Ada’s web import scraper now identifies itself with a static User-Agent: AdaWebCrawler header on every request. Websites that restrict crawling with an allowlist can add AdaWebCrawler to keep Ada’s scraper authorized — previously the user agent was dynamic and could not be reliably allowlisted.

Learn more in our Web import documentation.

Voice AI Agents can now play configurable audio cues while processing requests during phone calls, replacing silence with a subtle signal that the AI Agent is actively working.

What’s new

When your AI Agent needs extra time to process a request — such as retrieving information from external systems — callers previously heard silence after the initial acknowledgement message. This could lead to callers speaking again or hanging up, assuming the call had dropped.

You can now configure interstitial audio directly from the Voice settings in the Ada dashboard:

• Four sound options — Choose from Calm Contour, Bright Motif, Pure Cascade, or Slow Gallop. Each sound is optimized for phone audio clarity and designed to feel natural during a conversation. Preview each sound before saving.
• Adjustable delay threshold — Control how long the AI Agent waits (between 2.0 and 4.0 seconds) before interstitial audio begins playing.

To get started, go to Config > CHANNELS > Voice > Interstitial Audio in the Ada dashboard.

Learn more about configuring interstitial audio.

The Testing at Scale feature is now called Simulations. Dashboard labels, documentation headings, and URLs have been updated to match the new name. Existing links to Testing at Scale documentation redirect to the new paths.

For more details, see Simulations.

Customers with multi-brand Zendesk accounts can now complete OAuth setup when their AI Agent is associated with a brand-specific subdomain that differs from the primary account subdomain.

What’s new

Previously, OAuth authorization failed for multi-brand Zendesk accounts because brand subdomains are not valid OAuth endpoints. This blocked affected customers from completing the Zendesk OAuth migration.

A new Multi-brand toggle is now available in the Zendesk connection modal under Config > Apps > Zendesk. When enabled, you can specify both the brand subdomain and the primary account subdomain. OAuth authorizes through the account subdomain, and the connection is stored under the brand subdomain so that Handoff integrations find the correct credentials.

Learn more in Zendesk Handoffs.

Email AI Agents can now detect auto-reply messages that lack standard email headers, preventing reply loops that previously went undetected.

What’s new

Previously, auto-reply detection relied solely on email headers (Auto-Submitted, X-Autoreply, Precedence). Emails without these headers — such as some out-of-office replies, delivery confirmations, and automated messages — could trigger repeated AI Agent responses.

The AI Agent now uses AI-powered detection to identify auto-replies regardless of whether standard headers are present. When an auto-reply is detected:

• The AI Agent stops responding to the conversation
• The conversation closes automatically
• Auto-reply messages remain visible in the Conversations View — a “Conversation Resolution Analysis” event in the Coaching UI shows the detection result
• CSAT surveys are not sent for these conversations
• Webhook events continue to fire normally

Reply loops are now typically resolved within a few messages. No configuration is required.

Learn more about automatic reply handling.

We are releasing a new iOS SDK to align with Apple’s latest App Store submission requirements.

What’s New

• Built for the latest Apple toolchain – Compiled against the iOS 26 SDK, ready for App Store submissions under Apple’s new requirements.
• Modern installation options – Swift Package Manager is now the recommended install path, alongside continued support for CocoaPods, Carthage, and manual XCFramework distribution.
• Updated deployment target – Now targets iOS 16.0+, reflecting current device support. iOS 15 and below are not supported by this SDK.
• New MetaFields.Builder API – A cleaner, type-safe way to pass metadata and sensitive fields to your Ada bot at runtime.
• Minimal migration surface – AdaWebHost remains the main integration class, so for most apps the upgrade is a dependency swap and an import rename (AdaEmbedFramework → AdaMessaging).

To get started, see the iOS Getting Started guide.

Several improvements have been shipped to the MCP Server.

Write capabilities are now available via the new propose_change tool, enabling AI assistants to create, update, and delete knowledge articles, custom instructions, and test cases, as well as trigger test runs. The tool uses a two-phase flow — call first to discover available fields, then again to apply the change. Knowledge articles are created in a disabled state by default to prevent unintended live impact.

Containment metrics are now available in get_ada_metric via two new metric types: containment_rate (percentage of conversations handled without escalating to a human agent) and containment_volume (count of those conversations).

Conversation retrieval has been improved with a new get_conversations tool that returns many conversations at either an IDS_ONLY or SUMMARY detail level — reducing token usage compared to fetching full transcripts. Individual conversation lookup via get_conversation now also returns a conversation summary, resolution status, and resolution reason alongside the full event log.

Entity discovery is now handled by a single list_entities tool, replacing the previous list_channels and list_topics tools and adding support for discovering conversation variables (entity_type: 'variables').

Built-in prompts are now available in supported MCP clients. Two prompt templates ship with the MCP server: My Agent Report for a guided performance overview, and Export Testing for fetching and exporting test cases or test runs as CSV.

Learn more in our MCP Server documentation.

When your Voice AI Agent sends an SMS during a conversation, the full message content is now displayed directly in the conversation view.

This gives you complete visibility into the messages your AI Agent sends to end users via SMS, so you can review exactly what was communicated and how it contributed to resolving the conversation.

To learn more about reviewing conversations, see Conversations.

Playbooks can now search and reference Knowledge sources directly, without requiring AI Managers to copy article content into the Playbook manually.

New feature

Two methods are available:

• Direct Knowledge reference: Use @ArticleName in the Playbook editor to reference a specific article. Multiple articles can be referenced in a single step using comma separation.
• General Knowledge search: Include a natural-language instruction telling the Agent to search the Knowledge base at runtime. The Agent generates a search query based on the end user’s input and the instruction, then applies the returned content.

Both methods can be used together in the same Playbook.

The conversation view shows which articles were used—direct references display a Used article label, and general searches display a Referenced label with the articles retrieved.

For guidance on authoring Knowledge references, see Playbook management.