GuidesBy featureAnyone trying an agent that keeps working while they are away

Giving a persistent agent one job without handing over the keys

For anyone trying an agent that keeps working while they are away.

3 min read
7 steps
1 prompt to copy

The 2-minute version

Use it when You have access to an agent that keeps working after you close your laptop. It has a browser, a file system and a terminal of its own, and it can sign in to the tools you already use. You want to try it on something real without giving it the run of your accounts.

Carries out a repetitive job across several signed-in tools without being watched
Keeps working between sessions, so a routine can pick up where it left off
Holds context across a job that is too long to sit through, then hands you the result
Stops and asks before named categories of action when you have set those boundaries

In short

What this helps you do: Agents that run on their own cloud computer stay signed in to your tools between sessions. That persistence is the point, and it means the thing you are really configuring is access, not prompts.
What you leave with: One limited agent task, its permissions and the point where it must stop.
What to check before trusting it: You choose the one job that is safe to hand over, and the account it runs under

Ask Bit to set up this workflow

Tools to consider: Grok, Claude, ChatGPT or Gemini.Editorial starting points, not a tested ranking. Reviewed 19 September 2026. Open a tool guide for what each one does with this kind of work.

Workflow

0 of 7 done

Pick one job you could check in ten minutes and undo if it went wrong, and start there rather than with the impressive job
Create a separate account for the agent with access to only the apps and folders that one job needs, and do not reuse your own login
Sign in on the agent's machine yourself for anything needing a password, a two-factor code or a CAPTCHA, and understand that the session stays there afterwards
Write down what it should do when something is missing, so it flags the gap rather than filling it with last period's figure
Set approval boundaries by category before the first run: sending messages or invitations, publishing, purchases and transfers, deleting or overwriting, changing permissions, production changes, and accepting terms
Run it once, read what it actually did rather than only the output, and compare that against what you expected
When the trial ends, sign the machine out of every service and delete the files it left behind

Prompt to try

Use it, adapt it, and make it your own.

Prompt

Your job is [one specific recurring task]. Take the figures only from [named sources]. If a source has not refreshed or a number is missing, stop and tell me rather than estimating or reusing an older value. Before you send anything, publish anything, spend anything, change any permission or delete any file, stop and ask me first. When you finish, list every site you signed in to and every file you saved.

Ask Bit to adapt it

Human lens

You choose the one job that is safe to hand over, and the account it runs under
You type the passwords and the two-factor codes yourself, because those steps are human-only
You decide which actions must stop for approval rather than accepting the defaults
You sign the machine out and delete the files afterwards, because that part is not automatic

What to avoid

Separate agents are usually not separate accounts. Where they share one computer, one agent's logins, files and stored credentials are reachable by the others, whatever you named them
Deleting an agent tends to remove its profile, conversation and routines while files and logins on the shared computer may remain
Automatic review of an agent's own actions is itself a model making a judgement, and providers say plainly that it complements least privilege rather than replacing it
A broad allow rule written for convenience on day one is the rule that lets something through in week three, when the site it acts on has quietly changed
Persistence cuts both ways: an agent that keeps working also keeps holding access while nobody is watching it

Related guides

View all →

Handing a multi step job to an agentic model, safely

The newest flagships, Astra and Fable among them, are built to take on longer jobs that involve several steps rather than a single answer. That is powerful and worth using, but the more a model does on its own, the more it matters that you scope it, watch it and keep the reversible parts reversible.

Handing over a spreadsheet or file without handing over judgement

Modern assistants can read an uploaded file, whether a spreadsheet, a PDF or a plain document, and summarise it, clean it up or turn it into a chart. They are fast at spotting patterns and slow to notice when the data itself is wrong.

Letting an assistant take multiple steps on its own

Agentic modes plan a task, take a sequence of actions such as browsing, running code or editing files, and check their own progress along the way. They can save real time on chores with several steps, and they can also wander off course, so you set the boundaries and check the outcome.

Deciding whether a self hosted AI agent is worth the effort

Some agents run on your own infrastructure and remember across sessions, which buys control at the price of real upkeep. The choice of whether that trade is worth it, and the guardrails around it, stay firmly yours.

Reviewed 19 September 2026