The shift from assistants that answer to systems that act is arriving in ordinary business software, and the practical effect is narrower than the marketing.
What they actually do
Multi-step tasks with a clear definition of done. Extract data from a batch of documents and put it in a system. Draft responses to routine enquiries and queue them for approval. Reconcile records between two systems and flag mismatches.
Each of these was previously a person doing something repetitive and verifiable.
Why verification determines everything
An agent that completes a task wrongly, confidently, at scale, creates more work than it saves. The deployments that succeed are the ones where output is checked automatically or the error is immediately visible.
The deployments that fail are the ones where nobody would notice a wrong answer until later.
The permission question
An agent acting on your systems needs access, and access is where the risk sits. The pattern that works is read freely, write with logging, and require human confirmation for anything irreversible.
Organisations that skip this learn it expensively.
What it means for roles
Compression rather than elimination. The same output from fewer people, with the remaining work shifted towards checking and handling exceptions.
That is a real change in what the job is and it happens gradually enough that nobody announces it.
What to do if this is arriving in your workplace
Learn to evaluate output rather than resisting the tool. The person who can tell when the system is wrong becomes more valuable, not less.