2026-06-24 – 2026-06-30

Context Depth, which used to be available only on Enterprise plans, is now open to everyone. It controls how much information your agent has available before it answers - how much of your knowledge base it pulls into each reply, and how much of the ongoing conversation it keeps in mind.
- Available for every agent on the Intelligence tab.
- Pick from five presets - Compact, Balanced, Extended, Rich, or Max. More depth means more complete answers on complex questions and a longer memory across a conversation.
- A higher depth uses more query credits - with full visibility in the UI when setting it up.
- Enterprise plans additionally can get a fully Custom depth, set up at the plan level by their account manager.
Persona Expert Mode allows a more direct control over how the agent interprets its persona.
- Behind the scenes, every persona is now processed automatically each time you save it, turning your instructions into a set of helper notes that steer the agent as it works.
- Two helpers are generated: a short summary of the persona, and a routing map of which action the agent should use and when.
- On the Persona tab, Expert Mode allows power users to edit these helpers directly, separately from the main persona - useful when a complex agent needs to be told exactly how to handle a request without diluting your main instructions.
- Any change can be reverted at any time.
- Built with power users and advanced use cases in mind.

You can now set a custom "out of queries" message that logged-out users see when your account hits its query limit, instead of the default text.
YouTube playlists are now supported as a source, with automatic sync. Add a youtube.com/playlist URL and the agent reads every video in the playlist and keeps up as it changes.
Starting a new conversation on the Ask page now puts the cursor straight in the message box, so you can start typing right away.
PDF citations with limited page visibility can now be customized: rename the "Limited" label, and optionally turn it into a link that sends readers to the document's source URL.
When you upload images through File Upload without AI Vision turned on, a one-step prompt now offers to enable AI Vision so the agent can answer questions about what's in the images.
Enterprise Agents can now pull data straight from spreadsheet and tabular sources (csv, tsv, xlsx, ods) when searching, instead of falling back to plain text snippets.
The external API platform picker now shows more platforms per page, so you page through the list less to find the one you want.
The conversation history panel on agent deployments is now labeled "Chats" instead of "History".
More of the interface is now translated, including version history, the agent status messages ("Analyzing task...", "Reading your agent's persona...", and similar), and other strings that were still showing in English.

Enterprise Agents now support function calling through the Tasks API. You can give the agent a set of functions in the standard OpenAI tools format; when it needs one, the task pauses and returns the call, your system runs it and submits the result, and the task resumes until it finishes. This lets an agent read or change data in your own systems without an MCP server.
Deployment settings are now documented on the API. The PATCH project deployment-settings endpoint now specifies its request body and its section parameter, so you can update an agent's deployment settings programmatically.
POST /api/v1/projects/{id}/sources now auto-detects the source type from the URL, so you no longer need to send a type for YouTube or web pages. POST /api/v1/projects also accepts an is_master flag to create a Multi-Agent.
Enterprise Agent (Plan & Act) projects that call the OpenAI-compatible chat endpoint now receive a clear error pointing them to the /tasks endpoint, instead of a confusing failure partway through the response.

Fixed the issue where the browser address bar kept the previous conversation's link after you switched to an older conversation, so sharing or bookmarking it opened the wrong chat.
Fixed the issue where opening a cited PDF could fail with a permission error for guests on shared or public agents.
Fixed the issue where tools that were enabled but not actually used in a query - such as Document Analyst or a custom action - could still be counted toward your query cost. They are now billed only when used.
Fixed the issue where invited users had to pass through a blank email step when joining; the invite link now takes them straight to the account details screen.
Fixed the issue where feature announcements could send users without project-creation access to a "forbidden" page. Those users no longer see announcements they can't act on.
Fixed the issue where the Ask page showed a stray scrollbar even when the conversation fit the screen.
Fixed the issue where onboarding and billing pages scrolled into a strip of empty space below the content.