Use Atoms Agents to plan, build, research, analyze, and improve a project from its Chat. Turn Team Mode off for a focused request to Alex, or turn it on for Mike and the specialist team. This guide shows how to switch modes, mention Agents, follow questions and plans, queue follow-ups, review versions, and recover from an incorrect result.

How Agent teamwork works

With Team Mode off, Alex handles the request. Turn Team Mode on when the work needs Mike to coordinate several stages or when you want to mention a Product Manager, Architect, Data Analyst, or another available specialist.

In Team Mode, work can pass from one Agent to another. Follow Working Process and the Agent messages until no work is active. If the result creates a Version card, select Preview to open it in the App Viewer and test it before publishing or requesting another change.

Turn Team Mode on or off

The Team Mode switch is in the + menu beside the Chat input. When the switch is off, Alex handles the request. When it is on, Chat can use Mike and the available specialist Agents.

Before you switch: You cannot change Team Mode while Deep Research or Atoms Cloud is active. Deep Research uses Iris only, so the normal Group Members list is not available. Use the Agents shown in the current Chat. If Team Mode is fixed on and Alex is shown, mention Alex directly for an engineering task. To use the standard Team Mode choices, leave the dedicated workflow and open a normal project Chat.

1. Open the project Chat.
2. Select + beside the message box.
3. In the menu, turn Team Mode on for the team workflow. Turn it off when you want Alex to handle the request.
4. Close the menu, write the request, and send it.

If you type @ while Team Mode is off, Chat shows Switch to team mode to mention Agent. Select Team to turn Team Mode on and open Group Members, or select Engineer to close the picker and keep working with Alex without a mention.

Turn Team Mode on or off

Let Mike coordinate the work

When Team Mode is on and Mike is available under Group Members, send a message without an @ mention when you want Mike to decide which Agents should handle the work. This is the simplest option for an outcome that spans several stages. If Deep Research is active, use Iris instead.

1. Make sure Mike is available. Turn Team Mode on when the switch is available. If the switch is locked, use this workflow only when Mike appears under Group Members.
2. Describe the final outcome, the current state of the project, the important limits, and how you will judge the result.
3. Send the request without an @ mention so Mike can coordinate it. To address him explicitly, type @ and select Mike under Group Members.
4. Follow Working Process and the Agent messages. Answer a question when one appears. If a Plan card is ready, select Approve to continue or open its arrow menu and select Decline when the plan needs revision.

This kind of request gives Mike enough information to plan the work without requiring you to assign every step yourself.

Mention an Agent directly

Mention an Agent when you know who should receive a focused task.

1. In a normal project, select + beside the Chat input and turn Team Mode on. If a dedicated workflow locks the mode, skip this step and use only the Agents shown in the current Chat.
2. Type @ in the message box.
3. Under Group Members, select the Agent.
4. After the mention, explain the task, the important context, what must stay unchanged, and the result you expect.
5. Send the message.

Select an Agent from Group Members

Group Members normally includes Alex, Bob, David, Emma, and Mike. Sarah and Adrian appear in supported shadcn/ui or Atoms Cloud projects. When Deep Research is enabled, the mention picker shows Iris only. Use the list shown in your Chat as the source of truth.

Agent	Role	Use this Agent when
Mike	Team Leader	The task needs planning, delegation, or coordination across several Agents.
Emma	Product Manager	You need to clarify users, priorities, scope, workflows, or feature trade-offs.
Bob	Architect	You need help with system structure, data models, integrations, or technical constraints.
Alex	Engineer	You want to build, change, fix, or test the project implementation.
David	Data Analyst	You want to analyze data, files, metrics, or trends.
Iris	Deep Researcher	You need evidence-based market, competitor, audience, or technical research.
Sarah	SEO Specialist	You want to plan and improve SEO content for a supported web project.
Adrian	Ads Specialist	You want to prepare or analyze Google Ads for a supported published web project.

You cannot mention an Agent that is not shown under Group Members. Ask Mike to coordinate with the available team, or use the dedicated workflow that enables the specialist you need.

Work with more than one Agent

You can mention more than one Agent in the same Team Mode message. Use this for independent tasks that cannot overwrite one another, and state clearly which part each Agent should handle.

Separate each assignment clearly and state whether the Agent may change the project or should only review it.

For dependent work, do not send every stage to several Agents at once. Ask Mike to coordinate the order and include the point where you must approve the work.

Use only the Agents needed for the outcome, then review Credits used after the work is complete.

Follow tasks, approvals, and handoffs

Follow the work in the project Chat. Working Process shows active Agent steps, question cards wait for your answers, Plan cards wait for approval, and a Version card appears when the project produces a version you can preview.

What you see	What it means	What to do
Working Processshows an Agent as active	The current step is still running.	Wait for the result. If Queue is enabled, you can type the next instruction and press Return to queue it without stopping the active work.
A question card	Mike needs information before the workflow can continue.	Answer each question, useNextwhen there is another page, then selectSubmit.
A Plan card	The proposed plan is ready for review and execution has not started.	SelectApproveto continue. To request a revision, open the arrow beside Approve and selectDecline.
An Agent's final response	That Agent has finished its assigned step.	CheckWorking Processand later messages before treating the whole request as complete.
Another Agent starts working	The work has been handed to another specialist.	Follow the new Agent's messages and respond if it asks for information or approval.
A Version card	The project produced a version that can be reviewed.	For the latest version, selectPreview. For a historical version, selectLoad previewfirst and thenView Preview.

A written answer, Plan, or research summary does not by itself confirm that the project changed. When the request includes implementation, look for the resulting Version card, select Preview, and test the complete user journey in the App Viewer.

Send a useful follow-up

A useful follow-up tells the Agent what changed, what must stay the same, and how to check the result.

While an Agent is working, type the next instruction in the Chat input. When the input says Press Return to queue the message, press Return. The message is added to the Queue panel above the input and waits without interrupting the active work.

The Queue panel appears automatically after the first message is queued. Expand it if it is collapsed. Drag messages to reorder them; use Send now, Edit, or Remove on an individual message. A queued message with an attachment cannot be edited. The Queue header also provides Clear queue, Pause processing queued messages, and Resume processing queued messages.

If Queue is not available, wait for the current task to finish before sending the follow-up. Do not send the same instruction repeatedly, because duplicate requests can create conflicting work.

Review a queued follow-up

Recover when work is stuck or going in the wrong direction

Start with step 1, and move down the list only if the previous step does not solve the problem.

1. Check the current status. Read the latest Agent messages and Working Process. Look for an Agent that is still active, a question or Plan waiting for your input, or a follow-up waiting in Queue.
2. Correct the direction. Send one clear correction. Describe the intended outcome, the smallest change needed, and what must remain unchanged. Mention the Agent that should handle it when necessary.
3. Stop active work only when necessary. Select the square Stop control at the right of the Chat input. Its tooltip says Click if you want to stop all tasks. In the confirmation, select Yes, I'm Sure. This stops every active task in that Chat, and the interrupted response is marked Response stopped.
4. Restore a known-good version. If a completed change made the project worse, select Versions in the project header. Choose the last known-good version. If necessary, select Load preview and then View Preview. In the version preview, select Restore to this version. Before you confirm:
   • If Sync database appears, review it carefully. It is selected by default. Leave it selected if you want to restore the database with the version; clear it if you want to keep the current database.
   • Select Restore to confirm. The project returns to the selected version. Later versions remain available in Versions and can be reapplied if needed.
5. Report a problem with Atoms. If Atoms itself is not behaving as expected, open ... on the latest relevant Agent message and select Feedback. The current chat ID and message ID are attached automatically. Describe the exact steps, what you expected, what happened instead, and include a clear screenshot when useful.

Before you stop or restore. Do not use either action only because a task is taking longer than expected. Stopping can discard active progress. Restoring changes the current project state and can also change the database when Sync database is selected.

Stop all active tasks

Review the result, Credits, and sensitive information

Before you publish, update, or rely on the result, check that:

• Every requested stage is complete, including any handoff to another Agent.
• The final version opens correctly in the App Viewer.
• The main user journey works from beginning to end.
• The Agent kept the information, features, and design that you asked it not to change.
• No private information, credentials, or internal-only content appears in the result.

Review Credits used

On the last Agent message in a completed turn, open .... The menu shows Worked for; it shows Credits used after the turn is finished and usage is available. To check the workspace balance, open your profile menu and select Plans on desktop or Plans and credits on mobile. In the Plans and credits panel, review Credits remaining.

Review Credits used after each completed turn, and review the plan and scope before approving a large amount of work.

Review Worked for and Credits used

Do not paste passwords, API keys, authentication tokens, payment details, or unnecessary personal or confidential information into Chat.

FAQ

What should I do if Atoms refuses a request because it may be harmful?

Atoms may decline a request or ask for clarification when the requested result could be unsafe. Do not try to hide, disguise, or split up a harmful request to bypass that response.

1. If your goal is legitimate, explain the safe purpose, who authorized the work, and what the result must not do.
2. Remove real personal data, credentials, confidential information, and instructions that could harm another person or system.
3. Ask for a defensive explanation, safety checklist, or harmless example instead of instructions that enable misuse.
4. Review the revised request before sending it and keep it limited to the information needed for the safe task.

If a clearly legitimate, harmless request is still refused, contact Atoms Support with the Chat Link and a redacted copy of the request. Explain the intended use, but do not include sensitive information or ask Support to bypass a safety control.