Define the job before the model
Write a plain-language description of the work the agent will do. Specify the input, the expected output and the person accountable for reviewing its behavior. A bounded task is easier to evaluate than a general promise to assist the business.
Connect only the necessary context
Give the agent access to the documents and tools it needs for the task. Apply the same permission boundaries you would expect for a person in that role. Keep sensitive records out of retrieval unless access is explicitly required.
Distinguish drafting from acting
Drafting a response and sending it are different permissions. So are proposing a record update and committing it. Decide which actions can run automatically, which require approval and which should remain entirely outside the agent’s scope.
Evaluate uncertainty and failure
Use representative examples, ambiguous requests and deliberately incomplete inputs. Check whether the agent can identify that it lacks enough information. Track human corrections as evidence for changes to instructions, retrieval and permissions.
Start with one workflow. Understand it well. Build from there.
Talk through your workflow