> ## Documentation Index > Fetch the complete documentation index at: https://docs.dialtu.com/llms.txt > Use this file to discover all available pages before exploring further. # Atlas, testing and going live > The right panel: build with Atlas, test the draft, check the summary, activate The right panel of the editor has three tabs — **Atlas**, **Test** and **Summary** — and opens on Atlas. On narrower screens the three columns of the editor stack vertically. ## Atlas The first tab. Ask for changes in plain language and Atlas applies them to the agent. Atlas tab showing the request Make replies shorter, Atlas reporting Used 2 tools and confirming it tightened replies to 160 characters, an Edit SMS Agent Prompt receipt, and the two quick-action pills above the input What the editor tab can do: * **In create mode** it builds the agent from your description and drops you on its page. * **In edit mode** it updates the name, description, language, model, active state, agent type, first message, phone number, timezone, schedule, default-agent status (displacing the current default), auto close message, inactivity timeout, FAQ, variables, prompt text, attached tools, tags, disposition tags and memory settings. * The **Make replies shorter** and **Set inbound hours** pills prefill common requests. * It may ask clarifying questions before acting. What only the full Atlas chat does: delete agents, and create, update or delete tools — API, End Chat, Handoff and Send SMS only. Calendar and Integration tools are built in the UI. **Atlas writes to the saved agent immediately.** It's a publish, like pressing **Update Agent** — there's no preview step. **Atlas consumes credits.** If Atlas changes the agent while you have unsaved edits, a banner appears — *"Atlas updated this agent — your unsaved edits are still here."* — with **Reload** and **Keep my edits**. ## Test The second tab. You type as if you were the customer. Test tab reading Testing against your current draft, the two amber notices about real tools and credits, the Test as contact row, a customer message, a simulated Send SMS tool bubble and the agent's reply It runs against the **current draft** — unsaved edits included, even before the agent exists: name, prompt, FAQ, variables, language, model, auto-close message, memory settings and the selected tools. **API and integration tools run for real.** Use test data. **Each exchange consumes credits like a real conversation.** Unlike the WhatsApp test panel, this one is billed — at your plan's test rate, once, when you leave the editor or reset. See [What SMS costs](/en/sms-agents/billing). * **Test as contact** — pick a real contact to test with their memory and past conversations. The row then reads *"Memory: …"* or *"Nothing recorded about this contact yet."* It only appears if you can read clients and their notes. * **Send SMS** is simulated: no real text leaves. * When the agent ends or hands over, a pill marks the spot — *"The agent ended the conversation here."* or *"The agent handed the conversation to a human here."* — and the composer reads *"This conversation has ended — the agent has no way back from here. Reset to test again."* Press **Reset conversation** to start over. ## Summary The third tab. The checklist — **Required to create** or **Required to save** — lists **Name**, **Agent Prompt** (with its character count) and, for an outbound agent, **First Message**. Below it, a **Configuration** recap: Engine (Lumina), LLM Model, Language, Phone Number, Agent Type, Routing, Inactivity timeout, AI Memory, Status, Tools, FAQ and Disposition Tags. Summary tab with the Required to create checklist and the configuration rows, Engine reading Lumina ## Going live **Saving publishes immediately.** There's no server-side draft and no version history — *Draft saved* is a browser autosave. The control is the **Active / Paused** toggle in the top bar. A paused agent isn't picked for new inbound texts and can't start chats. **Duplicate** lives on the agent list, not in the editor. It asks for a **New Agent Name**, a **Phone Number** (required if the source is inbound) and, for inbound copies, **Default Agent** or **Time-Based Routing**. **The copy does not inherit the LLM model or the memory settings.** Set them again on the duplicate. **Delete**, also on the list, warns you: *"This action cannot be undone."* For big changes to a live agent, **duplicate** it, work on the copy, then swap which one is **Active**. ## What doesn't exist * **Version history** or rollback. * **Knowledge base** or document upload. * **Analytics inside the editor** — they're under **Analytics → SMS Statistics**. * **Bulk actions** on the agent list. * **Changing the engine.** ## Next step Where you watch it work — and step in.