The first task most people give an AI agent is too big and too vague, and the result teaches them the wrong lesson. Delegating to an agent works exactly like delegating to a capable new hire: a clear outcome, an explicit boundary, and a review you actually do. The difference is that with an agent the boundary can be enforced by the system rather than remembered by the person.
1. Write the Outcome, Not the Steps
An agent plans its own steps; that is the point of one. What it cannot invent is your definition of done. Tell it the outcome, who it is for, and what would make the result wrong. Triage the new requests into the right projects, assign by skill, and flag anything that mentions a deadline inside two weeks is a brief. Look at the backlog is not. In Kavanah a task assigned to an AI employee runs with the task's description as the brief and the project's charter as the standing context, so the brief is the task card itself.
2. Set the Boundary in the System, Not the Prompt
Here is the mistake the chat-first tools encourage: writing do not email anyone in the prompt and hoping. A prompt is a request; the model can misread it. A boundary belongs where the model cannot reach it. Give the employee only the scopes its job needs. Leave it in safe mode, which removes sends, payments and deletes from the run entirely. If the job needs a send, put an approval policy on external email so the draft waits for you. If there is one tool it must never touch, block that tool for that employee with a guardrail. Each of those survives a model upgrade and a rewritten prompt.
3. Give It a Negation
The KVN method asks every workspace to write down not only what it wants but what it refuses. A Negation is the list of things the organisation will not do, and in Kavanah the agent reads it before it creates work and declines what the charter forbids. This is the part teams skip, and it is the part that stops the agent from faithfully executing an ambiguity. An agent handed only positive direction fills the unspoken space with its own defaults, which are the model's, not yours.
4. Read the Ledger Before You Widen the Scope
Every action an AI employee takes lands in an action ledger: what it did, in which project, why, and whether it can be undone. For the first week, read it every morning the way you would read a new hire's work. Where the agent was right, leave it alone. Where it was wrong, resist fixing the output by hand. Write the rule down instead, as a Negation, a guardrail or a policy, so the mistake cannot recur. The ledger is also your undo button: a reversible action carries an inverse, so a wrong move is a keystroke, not an incident.
5. Widen in Steps
After a week of a clean ledger, add the next scope or move one category of action from approve to allow. Never both at once, because you want to know which change caused the next surprise. Most teams reach a point within a month where the standup, the triage and the status report run with no approvals at all, while client communication stays gated for good. That is not caution; that is the correct answer to how expensive a wrong action is.
The skill here is management, and most teams already have it. The agent just makes the boundary enforceable and the review cheap, which is more than can be said for the last three junior hires.
