Help

How to actually use the lab.

1. Give it something to do

Pick one of the three tabs above the input box — a simple one-line request, a real customer email someone actually wrote, or write your own message as if you were the customer — then press Run the agent.

2. Watch the transcript

"What it's doing" fills in turn by turn: the model's own text, each tool it calls with the arguments it chose, and the result that came back. A red entry means the tool call failed — the agent has to notice that and recover, same as it would with a real error.

3. Check its work

Everything a tool call changes — an order's status, an inventory count, a new ledger entry, an email in the outbox — shows up in the tables under "What's actually true," not just in the transcript. Newly changed rows flash briefly so you can spot what just moved.

Step through it one call at a time

Tick Approve every tool call before running. The agent will pause right before each tool call and wait for you to approve or deny it, instead of running straight through.

Start over

Reset the shop puts the orders, inventory, and ledger back to the same seed data, so you can try a different message against a clean slate. It's instant — nothing is actually persisted, so there's nothing to undo.

"No GeminiApiKey is configured"

This means live runs are switched off on this deployment. It isn't something you can fix from the page — it's a configuration flag the app checks before it will spend an API call.

← Back to the lab