Chat with Atoms to build your apps

Start tasks, layer on changes, and pick up right where you left off after an interruption. This page also covers every action available on a message, and what to do when an Issue Report appears.

Build with your agent team in chat

The chat panel is where you and your agents get things done. Say what you want in everyday language and the team handles the rest. Mike, your Team Leader, reads the request, turns it into a concrete brief, and passes it to the right specialist, such as Alex, the Engineer. The whole handoff happens in front of you, so you always know who is doing what.

Chat is not only for build requests. Ask why something works the way it does, what a term means, or what the team just changed, and you will get an answer right in the conversation.

Start a task

Type into the prompt box at the bottom of the chat and hit Enter. Focus on the outcome you want rather than how to implement it:

Create me a landing page to promote my medical equipment.

Once you send it, the work becomes visible immediately. Mike posts an activity card labeled “Processed N steps”. Expand it to read the activity narration step by step

Mike then writes a detailed brief: sections, constraints, and quality rules, then delegates it to the specialist with an @ mention. You will see the brief in full, along with the specialist accepting the handoff:

Mike assigns the landing-page build to Alex in Project chat.

Follow the work as it happens

Every agent reply carries an activity card. Expand it to see the activity narration for each step, plus tool cards for concrete actions such as reading a file, writing code, running a terminal command, generating images, or reviewing the rendered page.

Alex working through a build: each step is narrated, and tool cards like “Read file” and “Generate Images” show exactly what happened.

When a task wraps up, the closing reply summarizes what changed and how it was validated. A Version card, for example, “Version 1: Launch publishing hero” pins the delivered build, so every milestone in your project has a name you can point back to.

Mike’s completed build summary with the Version 1 card.

Add changes on the fly

Iterating does not require starting over. Send your next thought as a plain follow-up message: describe only the difference, and the team applies it on top of what already exists:

I need the theme to be green.

You can send a follow-up even while the agents are still busy. It will not interrupt anything: the message waits in a queue and runs as soon as the current task finishes. While a message sits in the queue, you stay in control of it:

• Edit it, if you want to sharpen the wording before it runs.
• Delete it, if you change your mind.
• Reorder the queue, if a later request should run first.

Version 1 has just been delivered — and a follow-up change is already on its way.

The follow-up goes through the same loop as any task: Mike converts it into a scoped brief, the specialist applies and validates it, and the result arrives as the next version:

The follow-up delivered: Version 2 applies the green theme while keeping everything else intact.

Interrupt a task and continue

Realized you forgot something important mid-build? You do not have to wait it out or throw the work away.

1. Stop the generation while the agent is working. Everything completed so far is kept.
2. Type what was missing: the detail, constraint, or correction you forgot.
3. Submit again. Mike folds the new information into an updated brief, and the team continues from the current state instead of rebuilding from zero.

Stopping Alex mid-build to add a forgotten requirement: “The equipment is mainly heart-related.”

After resubmitting, Mike issues an updated brief that reflects the new detail, and work resumes.

What you can do with each message

Messages in chat are not just text. A row of actions under each one lets you react without typing a word.

On agent messages

• Revert (back arrow): roll the project back to the state before that message’s changes. Confirm the target version first and review what may be removed before you use it.
• Copy: grab the message text, handy for saving an explanation or a summary.
• Helpful / Not helpful: rate a delivered result with a thumbs up or down to tell the team how it landed.
• More options (…): opens extra actions, including Feedback (the direct line to the support team when something goes wrong).
• Activity card: expand or collapse “Processed N steps” to inspect progress details and the tools behind the reply.
• Version card: the named build attached to a completed task.
• Identity label: every message shows who is speaking, Mike | Team Leader, Alex | Engineer, so multi-agent work stays easy to follow.

On your own messages

• They appear on the right side of the chat, keeping your requests visually separate from the team’s replies.
• You can send one at any moment: it runs immediately if the team is idle, or joins the queue if a task is in progress.

Issue Report: inspect and recover from problems

Problems can happen during building. The Issue Report helps you inspect the failure and choose a recovery path. When the agents hit a problem, an Issue Report notification appears in the bottom-left corner of the screen.

What the notification gives you

• A short description of what went wrong.
• A Resolve button that lets the AI attempt to analyze the reported issue and apply a fix.
• An expand control that opens the full Issue Report details in the right panel: error logs, stack traces, and system information, so you can understand the failure before acting.

You can select Resolve to attempt a fix, or read the Issue Report details first and then decide.

The Issue Report card: the blue Resolve button in the middle, and the expand control on the right.

Scenario 1: the page is stuck and there is no Resolve button

If the page freezes or stops responding and no Issue Report shows up, report it manually:

1. Open the chat with your agent.
2. Click the “…” (more options) on their latest message.
3. Choose Feedback, and include your Chat Link plus a screenshot of the problem.

This goes straight to the support team, who will follow up with you directly.

Scenario 2: you can see the Resolve button

Step 1 — Click Resolve

Select Resolve once. The AI agent attempts to analyze the reported errors and apply a fix. Wait for the attempt to finish, then review the result before continuing.

Click once, then wait. Clicking Resolve repeatedly before an attempt finishes spawns duplicate fix tasks and slows everything down.

Step 2 — Check the result

When the fix completes, the preview refreshes on its own. If it looks unchanged, refresh your browser first — stale cache is the usual culprit.

Step 3 — If the problem persists

Open the error logs in the right panel for the technical details. If the fix is beyond what you want to handle yourself, hand it to support:

1. Open the chat with your agent and click “…” → Feedback on the latest message.
2. Attach your Chat Link and a screenshot of the error.
3. The support team will investigate and work with you until it is solved.

Full reference: Issue Report — Atoms Help Center.

FAQ

Terms used in this FAQ: Preview is the in-editor view of the current project state. The latest version is the newest saved build; the last working version is the most recent build you confirmed was usable. The published site or live site is what visitors see after deployment. To get a Chat Link, open the relevant chat and copy its link from the chat’s sharing or more-options menu before contacting Support.

Can I send new requests while the team is still working?

Yes. New messages queue up and run in order once the current task finishes. Until a queued message runs, you can edit it, delete it, or move it earlier or later in the queue.

I stopped a task halfway, is the work lost?

No. Whatever was completed before you stopped is preserved. Add the missing detail, submit again, and the team continues from the current state.

How do I undo what a message changed?

Use the back-arrow action under that agent message to revert to the state before the change. Confirm the target version, review what will be removed, and then re-send your request with clearer direction.

I clicked Resolve but the error is still there. What now?

Check the error logs in the right panel first. If you are still stuck, use “…” → Feedback on the agent’s latest message and include your Chat Link and a screenshot, support will take it from there.

What if I close my browser mid-task?

The task may continue in the cloud while your browser is closed. When you return, check the task status and review the latest result; if you need to stop background work, reopen the project and select Stop when it is available.

What should I do when the AI agent stops responding?

Use this symptom-based recovery path; avoid sending the same prompt repeatedly:

1. If an Issue Report is visible, expand the Issue Report details and review the error before selecting Resolve once.
2. If the task is still running, select Stop once and wait for the status to settle.
3. If the page is frozen, refresh it or open the project in a new tab, then check the task status.
4. If output repeats or is wrong, review Preview and the latest version. Keep a usable result or use Remix from the last working version.
5. If a model error appears or several projects are affected, record the visible details and check the Atoms status page.

If the problem continues, contact Support with the Chat Link, last prompt, visible status, screenshot, time and timezone, and scope of impact. Remove passwords, tokens, and other secrets.

Why is the agent stuck in a loop or repeating itself?

1. Use the shared recovery path above: Stop an active run, review Preview and the latest version, then use Remix if the result is unusable.

If the loop returns after that path, contact Support with the Chat Link, repeated action, last prompt, time and timezone, and a screenshot. Do not include passwords, tokens, or secret values.

Why is the agent not following my instructions?

When the Agent ignores or misinterprets your instructions, try these four steps:

1. Refine your prompt — Be clear and specific. Use concrete criteria (e.g., "Change the primary button background color to #0055FF and center it on mobile views" rather than "Improve the button").

2. Use Design Mode — Switch to Design Mode and click directly on the UI component you want to modify. This ensures the Agent applies changes only to that specific element.

3. Break down complex tasks — Prompt one change at a time, verify the result, and then move on to the next requirement.

4. Provide expected vs. actual screenshots — Capture a screenshot of the actual result and share a reference image of the expected result. Visual comparisons give the Agent immediate clarity.

The agent says it finished, but nothing actually changed. What happened?

1. Open the latest version and compare the changed page or file with the requested result.
2. If Preview is correct but the published site is old, open Publish and select Update. Wait until the deployment is current, then hard-refresh the live site.
3. If only your browser appears stale, use a hard refresh or an Incognito window.
4. If no relevant change exists, retry with the exact page, element, expected result, protected areas, and acceptance criteria.

If the agent repeatedly reports success without a matching change, contact Support with the Chat Link, target page, expected and actual result, version number, and screenshots.

Why does the agent forget what I told it earlier in the conversation?

Long or mixed-purpose threads can make earlier requirements less influential. At a natural checkpoint, use Remix from the last working version and begin the new thread with a short handoff:

• Goal and current state
• Completed work
• Requirements that must remain true
• Files, pages, or screenshots that matter
• The single next task and its acceptance criteria

Verify the first result in the new thread before continuing. Keep unrelated work in separate threads so the agent does not have to reconcile competing instructions.

What are the current capabilities and limitations of the AI agent?

The agent can help create and revise application UI, implement scoped behavior, work with connected services, and investigate errors using the project context you provide. It works best when the target, constraints, and expected result are explicit.

The agent cannot guarantee that an unstated requirement will be inferred, that every generated change is correct, or that an external production service is healthy. It may also lose important context in a long or mixed-purpose thread.

For reliable results, give one verifiable task at a time, protect anything that must not change, review Preview and the latest version, test important flows, and confirm security, billing, and production-data changes before publishing.

The agent is not responding and I think the model is overloaded. What should I do?

A non-responsive agent does not by itself prove that the model is overloaded. Use the shared recovery path above; if several projects or models are affected, check the Atoms status page. Otherwise, review the result and retry one smaller request or use Remix from the last working version.

If the issue continues, contact Support with the Chat Link, model, last prompt summary, time and timezone, screenshot, and scope of impact. Remove confidential prompt data and secrets.

My task appears stuck and the page is frozen. How do I recover?

Use the shared recovery path above for a frozen page: refresh or open the project in a new tab, check the task status, and select Stop if the run remains active. If it stays frozen, contact Support with the full Chat Link, account email, screenshot, time of occurrence, and last successful message.

If the page is still frozen, provide the full Chat Link, account email, screenshot, time of occurrence, and the last successful message.

I'm seeing a model error. What does it mean?

1. Record the exact error text or code, selected model, task, project, and time. Then retry once or use Remix from the last working version; do not repeatedly submit the same request.

If the error continues, contact Support with the exact code, Chat Link, model, reproduction steps, and sanitized screenshot. Remove passwords, tokens, private data, and confidential prompt content.