Getting started with the Agent

Start with a bounded real task, inspect how the Agent works, and learn when to improve the setup rather than the prompt.

The fastest way to understand the Agent is to give it a real, bounded piece of work.

Do not start by trying to design the perfect prompt. Start with an outcome you already understand well enough to judge.

1. Pick a useful first task

Good first tasks have a clear result and accessible evidence.

Examples:

• compare a supplier confirmation with an open purchase order;

• extract and check information from a document;

• investigate a customer question using connected systems;

• prepare a structured summary from operational files;

• identify exceptions in a spreadsheet.

Avoid using your first attempt for a high-risk action where you cannot easily verify what happened.

2. Describe the outcome

Tell the Agent what you need, not only which feature to use.

For example:

Check this supplier confirmation against the purchase order and tell me what changed, what looks wrong, and what needs a person to decide.

Attach or point it to the information it needs.

3. Watch how it works

The useful part of a first session is not only the answer. Look at the work behind it.

Check:

• what context it used;

• which systems or tools it called;

• where it made assumptions;

• whether it surfaced uncertainty;

• what output or evidence it produced.

4. Correct the method, not just the answer

If the result is wrong, identify why.

Was information missing? Did the Agent use the wrong source? Is there an undocumented business rule? Should it have asked a person rather than deciding?

A one-off correction may only belong in the current session. A recurring correction should usually be captured somewhere durable, such as a Skill, maintained data, Knowledge Base or source system.

5. Turn repeated work into capability

When you find yourself giving the same instructions repeatedly, the work may be ready to become more reusable.

A Skill can capture a repeatable way of working. A Workflow can provide a more defined execution path where the sequence is known.

See Skills and Humans, Agents and Workflows.

A useful first milestone

You do not need full autonomy to get value.

A strong first milestone is often simpler: the Agent can perform useful work, expose its evidence, and reliably stop where a person still needs to decide.