Background agents: which ones actually keep working
Updated August 2026
Everyone ships one now, and the phrase has stopped meaning anything. Four questions separate an agent that works while you sleep from a chat window that stops the moment you look away — and you can put all four to any product in about twenty minutes.
Question 1 — where does it run when you are gone?
This is the whole category, and it is the question most marketing pages answer by not answering. There are three real shapes.
On your machine. The agent is a process on your laptop with a remote control attached — a chat bridge, a phone app, a web view. Everything works beautifully until the lid closes. Most “message your agent from anywhere” features are this shape, and they are not lying: you can indeed message it from anywhere, as long as the computer at home is awake.
On a machine that is yours to run. A VPS, a container, a home server. Genuinely persistent, and genuinely your problem at 3am.
On a machine the vendor runs. Nothing of yours is involved in it staying awake. This is the only shape where “it kept working overnight” is a property of the product rather than of your electricity supply.
Question 2 — what does it do at a decision point?
Every agent eventually reaches something it should not do unilaterally — overwrite a file, push a branch, spend money. There are only three behaviours, and the difference between them is the difference between a background agent and a foreground one wearing a costume.
It blocks and waits for you. Safe, and it quietly destroys the premise: an agent that stops after four minutes and waits ninety for you to reach your phone has not worked in the background, it has queued.
It proceeds under a policy you set in advance. Fast, and it means the policy is the product — read it carefully before you point it at anything you care about.
It carries on and gates the landing: the work happens unattended, and the moment where it becomes real waits for one tap. This is the shape that actually lets you close the laptop, and it is what we built.
Question 3 — does it remember, and on what disk?
An agent that starts from an empty machine every session is a very good session, repeated. The half-finished branch, the dependency it installed, the notes it left itself: either those survive until tomorrow or every long job restarts from nothing.
Ask two things. Is there a real disk that persists between sessions, or only a conversation history? And is that disk one agent’s, or one account’s — because if a fleet of agents shares a computer, they share everything installed on it, and one agent’s mess is another’s environment. At least one large vendor’s own documentation says this plainly, warning readers not to treat separate agents as a security boundary. It is a fair warning and worth taking literally.
Question 4 — how do you find out it finished?
The least discussed and most decisive one. If the answer is “open the dashboard and look”, the agent is not working in the background — you are polling it, and you will forget.
What you want is the same thing a colleague does: a message, in a place you already read, saying what happened. Not a notification that something completed — a sentence you can act on. The difference between “Task #4 finished” and “Checkout bug is fixed and pushed. It was rounding the total twice” is the difference between a tool and a hire.
The twenty-minute test
Whatever you are evaluating, including us, do this rather than reading the feature list.
Give it a job that takes longer than your patience — a dependency upgrade across a few files, a scraper, a migration. Then close the lid and go and do something else for half an hour, with your phone in your pocket.
Three things tell you everything when you come back. Did it keep going with your machine shut? Did it reach you, unprompted, when it got somewhere? And is what it did still there — the branch, the files, the installed tools — or did the environment evaporate with the session?
Anything that fails the first is a foreground agent with a chat interface. Anything that fails the third will keep passing the first two and never finish a job that takes two days.
What the answers cost
Broadly: agents that run on your own hardware are cheap in money and expensive in attention; vendor-run agents are the reverse. The per-seat products aimed at engineering teams are priced for teams, and the ones aimed at individuals are priced per month, usually with a usage allowance rather than a hard task count.
Whatever you compare, compare total monthly cost including the model — a cheap host with a bring-your-own-key model bill is not a cheap setup. That trap has its own page, because it catches almost everyone once.
Our answers, so you can hold us to the same four: it runs on a cloud machine of its own, one per agent, never shared with another customer or with your other agents. It carries on with the screen off. Anything that changes your code waits for a tap; anything that costs money is a checkout you complete yourself. And it messages you in Telegram when it is done. From $20 a month.
From $20 a month · cancel any time · ready in two minutes
Keep reading
- What Claude Code really costs to runSubscription or API, what the usage windows mean in practice, and the two bills people forget — the box it runs on, and the hours spent keeping it alive.
- Where to host OpenClaw so it stays upOwn hardware, a VPS, or a managed host: what each costs once you add the model bill nobody quotes — and an honest note about which reader we are not the answer for.
- Putting an AI agent in TelegramFour ways to get an agent you can message like a person — building the bot yourself, no-code wiring, Channels, or one bot per agent — and the gotchas that turn a weekend build into a month.