Building your first agent — bounded, read-only, supervised

The problem

You build an agent to research suppliers. You give it browsing and write access to your supplier sheet. It runs, follows a chain of links, spends forty minutes and a meaningful amount of money, writes eleven rows of partly-fabricated supplier data into your sheet, and you cannot tell which rows are real.

What you need

Goal: research a potential supplier before we contact them. Needs to look at: their website, MCA data if available, their marketplace listings if any, trade directories.

The tool

An agent mode in a major assistant, with tools restricted

The prompt

What comes back

1. GOAL, NARROWED — 'Produce a sourced one-page brief on [supplier name] covering: what they make, their scale indicators, how long they have operated, and any publicly visible customers or certifications.

Verify before you use it

• Confirm the agent has no write tool. You said the output is for a human, so it does not need one.
• Test the allowlist by giving it a supplier whose information is mostly on an aggregator site. It should report that and stop, not go there.
• Read all twenty traces. Watching only the output misses the reasoning failures that produce plausible briefs from almost nothing.
• Set the cost cap before the first run, not after the first surprising bill.