How to actually use an AI agent

Using an agent is closer to writing a runbook than to chatting with a genius.

1. Name the job

One job. "Screen these tickers for earnings quality" is a job. "Be my analyst" is not.

2. Name the tools

Which URLs, APIs, or files it may touch. If a tool is missing, it should say so and stop — not invent the filing.

3. Write the guardrails

What it must not do: buy/sell, invent numbers, impersonate a firm, write an exploit, bypass a site's rules. Put refusals in the prompt. Hope is not a control.

4. Keep a human check

The agent drafts. A person accepts. That is true for a 10-K brief, a code review, and a price comparison. If nobody reads the output, you did not delegate work; you added a log line.

Start with a short loop cap. Read one full run before you raise it.