Overview
Atlas is Voiceflow’s AI copilot for building and improving agents. Investigate customer conversations, propose instruction edits and create Tests from real cases to address recurring service issues. Atlas keeps one conversation as you move through the product, connecting investigation with editing.
Open Work with Atlas from your project sidebar. The composer’s edit-mode control switches between Manual and Auto. In Manual, Atlas asks before applying project changes; read-only work can proceed without approval.
Choose a starting point
- Why did a conversation go wrong? Ask Atlas to find and read relevant Transcripts, using a date range, Environment or Evaluation score to scope the investigation.
- Need to change agent guidance? Ask for an edit to the Global prompt, agent instructions or Playbook instructions. You can also select a passage in the editor and choose Ask Atlas.
- Want to preserve a conversation as a regression? Ask Atlas to create a Test from it, then run the saved Test yourself.
- Repeating the same request? Save it as a Skill in Settings → Skills and pick it from Atlas when needed.
Build and improve with Atlas
Refine the guidance behind an agent response. Review instruction changes in the native editor or ask Atlas about a selected passage.
Find the conversations behind a service issue. Search Transcripts, inspect Evaluation results and score a selected batch with named Evaluations.
Turn a customer interaction into a repeatable check. Review Atlas-authored Tests, Simulations and response, routing or tool-call checks, then start the run yourself.
Save recurring investigation or editing requests as Skills in Settings → Skills, then select them in Atlas.
Review and verify a change
Review the proposal
Choose Manual in the composer to review and approve project changes before Atlas applies them. Prompt edits appear as a diff you can accept or reject. You can continue editing the agent visually in Agent studio.
Complete connections in their own setup flow
If Atlas offers a Secret, integration or MCP setup card, open the normal connection flow and enter the credential there. The credential does not pass through the chat or model.
Run the verification
Atlas can author a Test, but you start its execution. Use Agent testing & QA to run the saved conversation in the intended Environment and inspect its results. For production trends and conversation evidence, continue in Agent monitoring & analytics.