Working with others
Agents: what they do alone and when they ask you
Work that advances without you, with a stopping point where a person decides.
An agent is work that advances without you present: it prepares, checks, sorts and completes whatever has clear rules. What makes it usable is not that it is fast — it is that it stops when a decision is due.
The states you will see
| State | What it means | What to do |
|---|---|---|
| Running | It is working | Nothing |
| Awaiting human | It reached something it does not decide | Look at it: that is where the person adds value |
| Pending review | It finished and wants confirmation | Review and approve or correct |
Important
"Awaiting human" is not a failure: it is the design. An agent that never stopped would be taking decisions that remain yours if something goes wrong. The more it stops early on, the more reliable it is when you let it run.
How to start without surprises
- 1
Review everything at first
For the first runs, approve by hand. That is where you see whether it gets it right.
- 2
Release what it always gets right
Repetitive work with a clear rule can stop being reviewed.
- 3
Do not release what decides
Approving, rejecting or blocking someone remains a person's job.
Watch out
It consumes credits, like any work using external services. If you leave one running over a large volume, check the balance first: it is the fastest way to exhaust it without noticing.
Worth knowing
Everything it does is logged like a person's actions: what it did, when and on what. If something does not add up, it can be reconstructed.
›Can I stop it?
Yes, and it is worth knowing how before you need to.
›What if it gets something wrong?
It is corrected like any other change, and the correction is recorded.
›Does it see more than I do?
No. It works inside your organisation and with its permissions.
A real case
The situation
A company leaves an agent preparing cases and reviews nothing for the first weeks.
What you do
- Goes back and approves the first batches by hand
- Releases only the repetitive part once it proves reliable
What you get
Ends up with 80% prepared automatically and the decisions with whoever answers for them.
The situation
An agent is switched on and the team does not know what it does unprompted and what it checks first.
What you do
- Checks which actions it can take alone and which ask for confirmation
- Starts with those that touch nothing outside
- Reviews after a week what it did before widening
What you get
The agent is widened on observed behaviour rather than on an expectation.
The situation
The agent writes to a supplier with nobody reviewing it.
What you do
- Keeps anything leaving the company behind a confirmation
What you get
Nothing reaches a third party unseen.
The situation
The agent approves a document that had expired.
What you do
- Limits what it can approve on its own
What you get
Approving the doubtful stays human.
The situation
Nobody knows what the agent did this week.
What you do
- Checks its action log
What you get
Its behaviour is auditable like a person's.
The situation
The agent gets something wrong and nobody knows what it decided on.
What you do
- Checks what information it had in front of it
What you get
The error is fixed in the criteria rather than case by case.
The situation
It is given access to every file for convenience.
What you do
- Limits its scope to the process it serves
What you get
The agent sees just enough for its task.
The situation
It is switched on and nobody decides when to switch it off.
What you do
- Sets a review at four weeks
What you get
The decision to keep it is taken with data.
This article answers
- what do ai agents do
- automated work that pauses to ask
- review what an agent did
- do agents consume credits