Atomic Agent

Atomic Agent: Phone and schedules

This page covers driving Atomic Agent from your phone and letting it work on a timer or overnight, then weighs each of those choices against what it puts at risk. It comes after Files, the browser, and memory, and you should not start it until your first jobs have worked.

Later: your phone and schedules

Wait until the first job has worked and you have looked at the folder yourself.

Telegram is off until you turn it on and add a bot token, which works like a password for your bot. Once set up, it answers only one owner: you.

The docs describe the setup in the terminal version's text screen:

Open it with atomic-agent, then go to Manage → Integrations → Telegram. Typing /telegram opens the same place.
In Telegram, message @BotFather, send /newbot, and follow its steps until it gives you a token.
Back in Atomic Agent, go to the Bot token field, press e, paste the token, and press Enter. The token shows in plain text while you type, so do not share your screen.
Press p. Your screen shows a 6-digit code; send /start <code> to your bot within 5 minutes, which makes you the owner. A message in a group chat never makes anyone the owner.

The older Telegram guide describes sending any message within 60 seconds instead. Follow whatever your screen shows.

When the agent wants a yes, your phone gets approve and deny buttons. If you ignore one, it turns into a no after 8 minutes, and you cannot change that time.

Your messages travel through Telegram's servers, and if you use a cloud model, they also go to that company.

The desktop 0.0.5 release notes describe Telegram pairing inside the desktop app, where Pair shows the one-time code. This guide could not confirm where that screen sits in the desktop app, or whether the desktop app keeps the bot answering on its own.

What you should see your bot answers you, and nobody else.

Discord

Discord is also off at the start, and it has no pairing step. Instead, you type your own Discord user number into the owner list. An empty owner list refuses every message, so put only your own number there.

Schedules and overnight runs

Atomic Agent can run jobs on a timer, every so often, or at set times, but there are two big warnings.

Scheduled tasks can skip your yes. The scheduling page says tasks run with the task command are "auto-approved," which means every action, including commands and file writes, runs without asking. Only schedule a job you would let run with no approval at all. This guide could not confirm whether a schedule started inside the desktop app asks first.
An overnight run in the terminal says no for you. atomic-agent overnight runs one long job with nobody watching, so anything that needs a yes is denied, and the end report lists it under "Needs you." Do not add --no-approval. In the desktop app, Run overnight is in the + menu next to the message box, and there a step that needs your yes waits for you in the app.

The computer must stay on, because schedules run only while Atomic Agent is running, and the desktop app must stay open during an overnight run. A missed run happens once, the next time it starts. This guide could not confirm whether schedules run while the computer is asleep, so plan on leaving it on, awake, and logged in.

If a repeating job fails too many times, it stops silently until you set it up again. The docs suggest having reports sent to Telegram for any job you depend on.

The main guide's phone and autopilot step explains these ideas with Agent Zero.

Trade-offs: safety versus capability

Approval level. The setting agent.approvalLevel runs from 1, the strictest and the default, up to 5. Most ordinary actions stop asking at level 4 or higher, and dangerous commands such as rm -rf and sudo keep asking until level 5.

You gain: a long job can run without stopping at every file write or command, so you can step away while it works.
You risk: anything that skips approval runs as you on the host, with no sandbox around it. A mistake or a prompt injection can then reach every file your account can reach.
Makes sense when: your account holds nothing you could not lose, which is almost never true for a beginner, so stay at level 1. This guide could not confirm a desktop screen that changes the level anyway.

Allow once, Allow in this chat, or Always allow. Your answer on an approval card decides how long a yes lasts. Allow in this chat covers that kind of action until the chat ends, while Always allow saves a rule that stays until you delete it in Settings › Privacy.

You gain: far fewer interruptions when a job repeats the same kind of step many times.
You risk: a yes you gave for one harmless step also covers later steps of the same kind that you never saw. A saved rule keeps applying in every future chat.
Makes sense when: you are watching a big, repetitive job on files you can restore, such as renaming a backed-up copy of photos in a folder made for that job. Keep Always allow for actions that cannot harm anything, and look over Settings › Privacy now and then.

Scheduled tasks or overnight runs. Tasks run with the task command are "auto-approved," so every command and file write goes ahead with nobody asked. An atomic-agent overnight run does the opposite and denies anything that needs a yes, listing it under "Needs you."

You gain: work that gets done while you sleep or are away, such as a daily report waiting for you in the morning.
You risk: an auto-approved task can run commands and change files on your real computer with nobody watching, and a mistake repeats on every run until you notice.
Makes sense when: not for a beginner, because a scheduled task runs every command with no approval on your real computer. Use an overnight run, which says no for you, and schedule a task only if you would let that exact job run with approvals off and no web reading.

Telegram on your phone. Telegram lets you send jobs and answer approvals with approve and deny buttons while you are away from the computer. It answers one owner, and an approval you ignore turns into a no after 8 minutes.

You gain: the agent stays useful when you are not at your desk, and you still get the final say on risky steps.
You risk: your messages pass through Telegram's servers, and a quick tap on a small screen makes it easy to approve something you did not read closely.
Makes sense when: your first jobs have worked, you are the bot's only owner, and the Discord owner list, if you use Discord, holds only your own number.

Local model or cloud model. The default is a local model through llama.cpp, while a free Gemini key sends your prompts to Google instead.

You gain: a local model costs nothing and keeps your prompts on your computer, while Gemini is stronger at jobs with many steps.
You risk: a local model on an 8 GB or 16 GB laptop tends to stall or loop. With Gemini, your prompts and any files the agent sends along leave your computer, and Google may use them to improve its products.
Makes sense when: you are starting first jobs on harmless files you do not mind sharing, which suits Gemini, or running short, simple jobs on private files, which suits a local model as long as you accept weaker results.

For the three questions behind this advice, see Safety versus capability.

Sources

Checked 14 November 2026.

Atomic Agent research notes.
Atomic Agent README (Discord owner list, overnight runs, Telegram pairing with a 6-digit code sent as /start <code> within 5 minutes).
Desktop v0.0.8 release notes (Run overnight, Always allow).
Release v0.6.11 (Telegram pairing with a 6-digit one-time code).
Desktop v0.0.5 release notes (Telegram pairing in the desktop app with Pair).
Approval gates (the five levels, dangerous commands).
Configuration (scheduled tasks at level 5).
Telegram guide (setup, one owner, 60-second pairing window, 8-minute auto-deny).
Scheduling (auto-approved tasks, runs only while the program runs, missed runs, failed jobs stop).
Overnight runs.
Security page (approval answers).

Next: Nine jobs to try.