Working with agents safely

An AI agent is not a chatbot. It takes actions: it writes files, calls services, and changes things. This page is about getting useful work out of one without being surprised by it.

Think of it as hiring, not prompting

When you create an agent you give it a name, a job title and a short list of things it is good at. That is deliberate. You are writing a very short job description, and the same rules apply as with a new colleague.

A vague brief produces vague work. "Help with marketing" gives an agent nothing to aim at. "Draft weekly customer emails from our product notes, in our house style" tells it what good looks like.

One agent, one job. It is tempting to make a single agent that does everything. Several focused agents are easier to judge, because when something goes wrong you can see which one did it.

Expertise is a hint, not a wall. The areas you list steer an agent towards the right kind of work. They do not stop it attempting other things.

Start supervised, earn the freedom

Every agent carries a trust mode, shown on its card:

Mode What it means
Supervised You approve its actions. Nothing happens without you.
Semi-auto Routine work runs; consequential actions wait for you.
Autonomous It gets on with its work and tells you afterwards.

Start new agents supervised. The first few tasks tell you more about how an agent handles your work than any amount of configuring. Watch what it actually does, then loosen the reins.

Move one step at a time. Supervised to semi-auto, and only later to autonomous. Jumping straight to autonomous means the first thing you learn about a mistake is that it already happened.

Autonomy is a promotion, not a setting. Give it for demonstrated good work on a particular kind of task, the way you would with a person. If an agent starts producing work you have to correct, move it back down. That is not a failure, it is the system working.

What GNexusOS will never let an agent do alone

Some things are not left to a trust setting. The app has a built-in constitution that applies no matter how much freedom you have given an agent.

These always stop and ask you, even for an autonomous agent:

  • Destructive actions on things outside your own machine
  • Anything involving money
  • Creating public accounts in your name
  • Copying or replicating the platform itself

This is refused outright, and you cannot approve it:

  • Running anything with system administrator privileges

That last one has no override. There is no setting that enables it, which is the point.

And some places are permanently off limits to agents that can control your screen: banking sites, payment processors, brokerages, crypto services and password managers. This is not a default you can change.

Why it works this way. A setting you can turn off is not a safeguard, it is a suggestion. These limits hold regardless of how you have configured anything, so a mistake in your settings cannot remove them.

Approvals are where you do your real supervising

When an agent needs permission, the request appears under Approvals.

Read what it is actually asking. The request says what the agent intends to do. Skimming and approving defeats the purpose of having asked.

Rejecting is normal and useful. It is information, not a rebuke. The system records your decisions and they build up a picture of what you are comfortable with.

Do not let approvals pile up. A long queue leads to approving in bulk to clear it, which is how an unwanted action slips through. Keep it short enough that you actually read each one.

Give agents the context, not the credentials

Point agents at your workspace. Everything in it — notes, files, past work — is available to them, and an agent with the real material does better work than one working from a description of it.

Never paste an API key, password or card number into a task. Credentials go in Settings, where they are stored in your operating system's secure storage. Anything typed into a task becomes part of the record of that task.

Be deliberate about web search. It is a per-agent setting. An agent that can search is more useful and less predictable. Turn it on where it earns its place.

Keep an eye on the record

Check the awareness feed. The right-hand side of the main window shows what has been happening. A quick glance tells you whether anything unexpected ran.

Flags are worth reading. When the system notices something that does not add up — work claimed as done with nothing to show for it, for example — it raises a flag. Those are the most valuable thing on the screen, because they are the cases where an agent's account of itself and reality disagree.

Your memory improves with use. Finished work is turned into organised notes automatically. The more an agent does, the more context later work has.

A sensible way to start

  1. Create one agent for one real job you actually have.
  2. Leave it supervised.
  3. Give it three or four real tasks and read everything it produces.
  4. Move it to semi-auto once you stop finding things to correct.
  5. Only then add a second agent.

The temptation is to build a team of eight on the first afternoon. A single agent you trust is worth more than eight you have to check.