Changelog

Open an app beside the conversation

Apps now open in a panel next to the chat, so you can follow the thread and use the app at the same time. Ask the Agent for a change and the app updates in place. Your filters, selections and anything you’d typed into it stay where they were.

Switch Dema to dark mode

Dema has a dark theme, and it is on for every account. The app follows your device by default, so it turns dark when your laptop does. To choose for yourself, open Settings > Account > Profile, and pick System, Light, or Dark under Appearance.

Dictate a message instead of typing it

You can now speak to the Agent. A mic button sits in the conversation input, and what you say appears as muted text trailing whatever you had already typed, with a waveform underneath so you can tell the microphone is picking you up. It understands English, Swedish, Norwegian, Spanish, Italian and French, and it knows Dema's own vocabulary, so terms like MMM and CFA come out right.

The first use asks your browser for microphone permission, and where the browser or your operating system has it blocked, the button tells you which setting to change.

See what your competitors are running

Your agents can now search the Meta and TikTok ad libraries. Point one at a competitor, and it can pull the ads they are running, then compare those angles against your own inventory. You can also ask an agent to keep a running ledger of what each competitor tries, so new angles show up as they appear. Coverage is limited to ads shown in the EU.

Run incrementality tests across countries

Incrementality experiments are no longer limited to one treatment country. When you create an experiment, you can now pick several. Country tests can give cleaner results than region tests in certain cases. Ads don’t leak across borders the way they do between regions, and national sales volume makes lift easier to see. The results line up with how you already budget and report. They also work in markets where one city dominates, where a region test can’t find a fair control group.

Build Meta campaigns from the conversation

Agents using the Meta Ads integration can now create and update campaigns, ad sets and ads, and upload the images for your creatives, so a campaign can go from brief to built without leaving the chat. New campaigns, ad sets, and ads are created paused. Nothing goes live until it is switched on, and you can keep the integration's write tools on ‘Needs approval’ if you want to see each change first.

Your schedules have a home of their own

Scheduling an agent used to sit behind a button on the agent page. That button now takes you to Routines, a page where your scheduled work lives together.

A routine runs a task automatically on a daily or weekly schedule. Use routines for work that should happen consistently, such as reviewing stock every morning, updating a performance app each day, or preparing a channel summary every Monday. The full walkthrough is in the routines guide.

Hand an agent to someone else

Agents and apps can now change owner. Click the owner's name on either one, pick a colleague from the dropdown, and it's theirs.

The useful part is what travels with it. Agents rarely stand alone: they carry skills and apps, and some of those are shared with other agents. The transfer dialog lists everything linked to the agent and lets you decide what moves. Skills and apps that only this agent uses are ticked already, so for most handovers you can accept what's there and confirm. For an agent that shares its skills more widely, you get to choose who ends up owning what.

Tell each integration what it may read and what it may change

An agent's integrations now separate the tools that read your data from the tools that change it. Open a connection, and you get Read tools and Write tools as two sections, each with its own count and permission summary, so you can tell at a glance whether something can only look at your ad accounts or can also spend from them.

Each section has a dropdown that scopes every tool inside it at once: Always approved, Needs approval, or Blocked. Individual tools can still differ, and the section then reads Custom. Tool names are now also spelled out in plain language.

See how each app stays up to date

You can now see exactly how an app gets fresh data. Open any app, and the Updates tab names what powers it, the agent or skill behind it, along with when its data was last refreshed and when the next refresh is due. What used to be Runs is now History.

Update now appears only when it will actually do something. For an app driven by an agent, choosing it opens the conversation that does the work. For an app that fetches live data itself, the figures refresh in place, and you stay where you are. Where an update isn't possible, the tab explains why rather than leaving you with a greyed-out button.

Which updater an app uses is the decision of the agent that built it, since it depends on what that app's data needs. If you want it changed, ask for it in the conversation.

Two more integrations

Microsoft Business Central joins the agent integration catalogue. Microsoft's ERP covers accounting, inventory, orders, and fulfilment, so your agent can reason about stock and financials alongside the marketing data it already sees. Access is read-only. Setup is in the Business Central guide.

Klaviyo now sets up like any other connection. Add it from Integrations under Settings, enter a private API key, and the segment export runs from there. It follows Voyado, which became self-serve in August.

A conversation you can actually read

Agent conversations now render at a reading size rather than the size the rest of the interface uses for dense controls. Answers and your own messages step up from 13 to 15 pixels, and headings follow a ramp built for prose, so a heading never comes out smaller than the paragraph it introduces.

Code is more consistent as well. A tool card used to mix two monospace typefaces across three sizes, and input, output, and diffs now share one. Inline code scales with the surrounding text, while fenced blocks stay fixed, since a block is its own surface rather than part of a sentence.

The right-hand panel got the same attention. Section headings collapse now and remember what you left open.

Your agents keep a history now

Every edit to an agent, an app, a skill, or an instruction is now saved as a version, and you can restore any of them.

Open the history on any of the four, and you get the full list, newest first, with who made each change, when, and the note they left. Pick one and revert. The revert is saved as a new version, so nothing is lost on the way back, and you can always undo the undo.

Files come with it. A skill or agent's uploaded files are now kept per version instead of being overwritten on each edit, so restoring an older version brings back the files as they were at the time, not whichever ones happen to be there today.

Reverting needs the same permission as editing. If you can change it, you can restore it.

The Integrations page tells you which storefront

The Integrations page under Settings has been redesigned around the question people kept having to ask us: which storefront is this actually connected to?

Each row now shows how many of its connections are live out of how many exist, and expands to reveal the accounts behind it. Click a row and a panel opens with that integration's connection details and setup instructions. It's also where you connect a new one, so Add data connection is no longer the route for anything already in the list.

Seven more sources joined the catalogue while we were in there: BeX, BigQuery, Databricks, Google Sheets, Salesforce, Sitoo, and Snowflake.

Two Shopify stores, one agent

If you run more than one store, region, or brand on the same platform, you can now connect all of them to a single custom agent instead of building one agent per account.

Add each connection from the agent's Configuration tab, and the agent addresses them separately at runtime, so one agent can pull from your US and EU stores in the same conversation. This works for every integration that supports multiple accounts, including Shopify, Klaviyo, Centra, NetSuite, Snowflake, Storyblok, and any custom MCP server you've added.

Shared apps work for the people you shared them with

Share an app with your team, and it now runs for them, including when the agent behind it stays private to you. They get the app and its data without needing access to the agent that produces it. Previously, the app would only run for people who could also see that agent, so a private agent meant a shared app worked for you alone.

Approvals are unchanged: each tool still resolves against whoever is using the app, so nothing runs without the permission it needs.