Agents are the wrong tool more often than their advocates admit, and the cost of finding out late is a confident mistake in production. Three questions filter most bad delegations before you start.
1. Can you describe the goal without describing the steps?
If the task requires you to specify every action, you want a script. An agent's value is choosing the steps; if there is only one possible sequence, you are paying for improvisation you do not want.
Corollary: tasks with a fixed, boring procedure should be automated with code, not delegated to a model. A shell script is deterministic, free, and does not hallucinate.
2. Can you verify the result cheaply?
This is the question that decides most cases. Running tests costs seconds. Reading a diff costs a minute. Checking a legal paragraph against a regulation costs an afternoon — and that afternoon is the real price of the delegation.
If verification costs more than doing the task, the agent is not saving you anything. It is moving the work from doing to checking, which feels faster and usually is not.
3. Is the action reversible?
Reversible: drafting, summarising, proposing a diff, querying a database with a read-only role. Irreversible: sending, publishing, deploying, deleting, moving money, rotating credentials.
For irreversible actions, an agent can prepare but a human must commit. This is not distrust of the model; it is the same rule you apply to a new hire, and for the same reason.
The signature of a bad delegation
It usually looks like this: the task is vague, the verification is "I will read it carefully", and the action touches something you cannot undo. Each of those is survivable alone. Together they are how teams end up explaining an incident.
Where agents are clearly right
Tedious work with an obvious check: renaming across files and running the tests, writing the test that reproduces a bug, reading six files to find where a config is set. Boring, bounded, verifiable.
For the tooling side — what to connect, and how much access to give it — read guardrails.
Next: back to the curriculum.