Review Atlas prompt and instructions changes before approving

Before you approve an Atlas change to your global prompt, your agent’s instructions or a playbook’s instructions, you can see exactly what changes, with the stored text set against Atlas’s proposal.

Click Review changes on the approval strip above the Atlas composer.
Read the stored text against Atlas’s proposal in the changes review.
Click Approve or Reject on the strip.

This applies when Atlas is in Manual mode.

Read and copy chunk metadata in the knowledge base

Hovering the grey metadata line above a chunk in a data source now shows a Metadata readout, and a Copy as JSON button in the tooltip copies it in one click.

Each metadata key and its value get their own row, up to 10 rows, followed by an and N more count.
The smart-chunking options applied to the chunk are listed below the rows.

Find it under Knowledge base > Data source > Chunks.

Messages sent mid-turn interrupt your agent when possible

A message that arrives while your agent is busy, for example during a tool call, is now held on v4 socket and voice conversations instead of being rejected. It interrupts the turn at the first point where the turn can be interrupted, or runs when the turn ends.

Several messages sent while one is waiting are merged into a single turn.
A newer non-message action replaces an older held one, and the older one is rejected as superseded.
A held end is never replaced.

Read evaluation release cards in full over the criteria table

A release card now opens in full over the criteria table when you hover a release marker on an evaluation card on the Evaluations page.

Close an evaluation’s log by clicking its row again

You can now close an Evaluation log by clicking its open row again in the Results table of an evaluation.

Clicking a different row still switches the log to that row.

Agents can link knowledge base source pages

Your agent can now link the page an answer came from: each result it reads from the knowledge base includes that page’s address, such as a web page URL or a Zendesk article’s public help center link. This holds even when Show source URL(s) is off.

To have the agent link the page it used, ask it to in its instructions.
To keep it from sharing links, tell it not to in the same place.
Show source URL(s) still adds its separate citations in chat projects.

This works for agents that have the knowledge base tool turned on.

Shopify imports now include only active products

Importing a Shopify store as a data source now adds only active products to your knowledge base and the Add data sources count, not draft or archived ones.

Atlas can now optimize long playbooks and Main Agent instructions

Atlas can now shorten a long playbook or the Main Agent’s instructions without losing a rule, or move a separate job into its own playbook with its tools, exit paths, routing and tests.

Ask Atlas to optimize, shorten, clean up or split a playbook or the Main Agent’s instructions.
Accept the suggestion Atlas offers once after it reads instructions of 25,000 characters or more.

Atlas reads the whole text and everything linked to it, then writes one plan with the current size and the target size and asks once for your go-ahead in the conversation. After you approve, it runs the tests the plan names, makes every change, runs them again and reports the passed checks before and after.

Voiceflow Classify

Voiceflow Classify is a new intent classification option for V3 projects. It uses Voiceflow’s decision model to pick the intent that best matches the user’s message. It’s faster than classifying with a large language model and needs no training. V4 projects don’t have intents, so Voiceflow Classify doesn’t apply to them.

There’s no confidence threshold to tune: when no intent clearly fits, or the message is too vague to tell, Voiceflow Classify returns no match instead of guessing. If Voiceflow Classify can’t classify a message, for example because more than 254 intents are in scope, your agent falls back to Large language models (RAG + LLM) for that message.

To switch, open Intents, select the settings icon next to Preview, and choose Voiceflow Classify. Projects keep their current setting until you change it.

Give each intent a few clear utterances. Descriptions are optional, but they help separate intents that are close in meaning.

Atlas can now give your agent a phone number

Atlas can set up a phone number for your agent from a Phone number card in its chat. Ask Atlas to set up a number, or let it offer one after it builds or publishes a voice agent that has none.

Create number and Add existing number open the same forms as the Phone numbers page.
When you finish, the card tells Atlas which number was added.

When Atlas moves, disconnects or deletes a number that answers for an agent, an approval card names that agent and waits for you, even in Auto.

Agents reply after receiving a conversation from another agent

When one agent or playbook passes a conversation to another, the agent that receives it now replies instead of going silent, even if the history mentions tools it cannot call.

Filter the knowledge base table by source type

You can now narrow your knowledge base to the kinds of source you care about, such as only sitemaps or only Zendesk articles.

Click the All data sources header of the table and tick the kinds to show: Web pages, Sitemaps, Docs, Tables, Plain text, Zendesk or Shopify.
The header then names the kinds shown and counts them.
Unticking every box shows all sources again.

If nothing matches your filter, the empty table offers Clear filters.

Knowledge base table sorts by last successful sync

The Synced column in the knowledge base table now sorts web page and sitemap data sources by their last successful sync, not their last edit, and the status checkmark shows the exact time.

Atlas suggests tightening long playbook instructions

Atlas now offers to tighten or split a playbook with a tip on its launcher once the playbook’s instructions pass 25,000 characters.

In Agent > Playbooks, open a playbook and stop typing for a moment, and a card on the Atlas launcher says how long the playbook is.
Click Optimize with Atlas to open Atlas with a request to find what is repeated, contradictory or out of date, suggest splits, and show the changes before making them.
Click Not now to hide the tip until the playbook grows by a quarter or two weeks pass.

After two Not now answers, the tip stops appearing for you.

Set savings as one dollar value per resolution

The Estimated savings dialog now asks for a single number: the dollars a resolved conversation is worth. On Analytics, choose Define savings (or Edit) on the Resolution card and enter it in Value per resolution.

The agent’s Objective settings and Atlas home ask for the same value.
The Time saved per resolution field is removed from the dialog.
The Resolution rate and Deflection rate evaluations no longer show Time saved (hours) and Dollars saved with Add estimate in the evaluation editor.

Attach files to Atlas messages

You can now attach documents, spreadsheets and screenshots to Atlas messages, in the Atlas panel or on Copilot Home, and ask Atlas to build or answer from them.

Attach up to five files per message with the attach button in the composer, by dropping them in, or by pasting.
PDF, Word, Excel, CSV, text, Markdown, PNG, JPEG, GIF and WebP files are supported, with documents up to 10 MB and images up to 5 MB.
Each file shows a chip that reads Reading, then ready, or failed with a reason.
Atlas reads each file as text, and reads further into a long file when it needs to.
Once you approve, Atlas can add a file to your knowledge base.

See what each Atlas write changed, step by step

A finished Atlas write now has View changes beside its row, opening a modal that shows each change in the reply as Before and After cards, so you can check the result without leaving the chat.

A collapsed run shows View N changes in place of the single button.
The modal steps through every change, such as Change 3 of 7, with Previous and Next or the arrow keys.
A change made in another environment or agent has Open in that environment, which switches to it and opens the item.

Changes too large to keep say so instead of showing cards.

Keep Zendesk structure when importing a help center

A Zendesk import into your knowledge base can now keep the help center’s structure, filing articles in folders by language, category and section, the way Zendesk shows them.

In Knowledge base, choose Add data source, then Zendesk.
Under the folder picker, tick Preserve Zendesk structure.

Articles are filed inside the folder you selected, under a folder per language such as English (en-us), then its category, then its section and any nested sections. Importing again reuses the same folders.

Knowledge base answers cite Zendesk help center pages

Your agent can now cite the public help center page of an article from a Zendesk data source as a source.

With Show source URL(s) on in the knowledge base tool settings of a chat agent, a reply based on a Zendesk article lists that article’s help center page as a source.
In the knowledge base, opening a Zendesk article or a Shopify product shows a read-only Source URL that opens the public page in a new tab.

Articles you imported before this change get a Source URL too.

Keys need new permissions for protected environments

You can now give a personal key or service key permission to merge into, publish and delete protected environments, or keep it from doing so. A key created with chosen permissions is now refused on those actions unless it carries the matching permission.

In Workspace Settings > Personal keys (or Service keys) > New key, choose Custom under With these permissions to find the Protected environments group with Merge, Publish and Delete.
Read & write now includes Merge and Publish.

Keys created before this change with chosen permissions don’t carry the new ones and must be recreated. Keys with no permission limits, and AI clients signed in to the MCP server with full workspace access, are unaffected.

Secrets in handoff steps

A Live agent handoff step can now take its value from a project secret, not just a variable. This lets you keep values like an integration ID or a queue ID out of plain view while still populating the step.

In the handoff step’s fields, choose a secret anywhere you could already choose a variable. The secret is resolved when the handoff starts.

The step’s Configuration JSON stays variables-only and does not accept secrets.

Framework settings

The Framework page no longer has its own place in the settings rail. The initialization workflow now lives under Behaviour, alongside the rest of an agent’s behaviour settings.

If a project is still on conversation flow, you can switch it to the agentic framework from the General tab. The switch only goes one way: once a project moves to agentic, it cannot move back.

This only applies to projects still on conversation flow. An agentic project has no framework switch to make.

Global prompt timezone

Your agent’s timezone can now be set from the global prompt editor, right next to the prompt that uses it. Your own timezone is offered first in the list, so the common choice is right there.

Open the Global Prompt tab and use the timezone selector to choose a zone, or clear it if you no longer want one set. Clearing now removes the value instead of requiring you to replace it with another.

Transcript version filter

The Search transcripts endpoint on the Analytics API can now search across several versions in one request.

Pass a list of version IDs to the versionIDs field to pull transcripts from a published version and its past releases together, instead of querying each version separately.

Messages and prompts API

Messages and prompts can now be read, created, edited and deleted by ID through the API, the MCP server and Atlas. This includes a message’s variants and a prompt’s system text and settings, so scripted content can be managed directly rather than only through the builder.

Use the API or MCP server endpoints for messages and prompts to fetch or update them by ID, or work with them in Atlas. The /v2 environment export now nests these under messages and prompts, so exports reflect the same structure.

Changelog