Will the agent replace our team? No. It takes over the repetitive steps: the lookups, the copying between systems, the first draft. People keep the decisions and the cases that need judgement, and they get them with the facts already collected.
Which agent framework do you use? Often none. Many agents are clearest as plain code around a model’s tool calling. When a framework earns its place, for example for long tasks that must survive a restart, we use one and explain why.
Our systems have no API. Can an agent still use them? Sometimes. It can read a database, work with exported files, or call a small API we build in front of the old system. We avoid agents that click through screens, because they break when the screen changes.
How do we stop it? With one switch for the whole agent and one per tool. Turning it off sends new work back to your team, and nothing in progress is left half done.
What does each task cost? We estimate it before we build, from the steps a typical task takes. After launch, the cost of every task is in its log, and limits stop any task that runs over.
Is it safe to give it write access? Only after it has earned it. An agent starts in shadow mode, moves to drafts that a person approves, and acts alone only on the low-risk cases you choose.