Put an agent in your team's hands

Publishing an Agent takes one click. Getting a team to use it involves three things the publishing screen doesn't mention: who can access it, how much it will cost and how you'll know whether it's working.

This guide covers those three.

Before publishing

A short check prevents most first-day problems:

• The documents are Active. Collections still indexing make the Agent answer as if it were empty.
• The model is still enabled in the workspace. A removed model breaks the Agent without warning.
• A real question was tested, and the cited reference supports the answer.

Who will be able to use it

Chatting in Chat only requires being a workspace member. No administrative role is needed, and User is enough.

Adding someone takes two steps, on different screens:

1. The person must belong to the organization. If they don't yet, someone with the Admin role invites them.
2. The person must be added to the workspace, by its Owner.

Estimate and cap the cost

Before opening it to many people, it is worth having a sense of the order of magnitude.

Let the Agent run with a small group for a few days and watch the per-model consumption in Observability, and the per-Agent usage in the Dashboard. From there:

• Set a Workspace cap with headroom over the observed peak, not over the average.
• Consider a per-User cap, which is what protects against a single case eating the whole month.

Starting too tight causes blocks on the first day of heavy use, and the fix ends up becoming a permanently high policy.

Communicate what it does

The Agent's name and description are the only instructions most people will read. They show up in the Chat selector.

Describe the job and the limit: "HR policy questions: vacation, expenses and benefits" says more than "HR Assistant", and reduces out-of-scope questions.

The first weeks

Three screens, in order of usefulness:

• Logs › Response feedback: where problems show up first. Low volume is normal; people rarely rate answers, and almost always when something went wrong.
• Logs › Conversations: what people actually ask. It usually yields the second version of the prompt and a list of missing documents.
• Dashboard: whether usage holds up once the novelty wears off.

Changing an agent in use

There is no draft of an already published Agent. Saving a change puts it into production on everyone's next conversation.

For large changes, duplicate the Agent, adjust the copy as a draft and only then swap it in.

Next steps

• Adjust the instructions based on what you read in Write the prompt.
• Set the caps in Set limits.