Your first agent that actually does something

The trap to avoid

The most common failed first agent is a general assistant with access to everything and a vague brief. It demos well and then nobody trusts it with real work. The fix is to go narrow: one job, one table, one example you can check by hand.

Build it

1. 1
   Pick one judgment-heavy step your team repeats — labeling inbound requests is the classic. Write down three real examples and what a good output looks like for each.
2. 2
   Create the table first (Data → new table). Give it the columns your output needs: title, category, priority, status with a default of OPEN.
3. 3
   Create the agent (Agents → New agent). Instructions are a job description: what arrives, what to produce, what to do when unsure. Paste one of your real examples into the instructions as a worked sample.
4. 4
   Open the Access section and grant exactly one thing: write access to your new table. Resist granting more.
5. 5
   Run it from chat with your second real example. Open the conversation and expand the tool calls — check the row it wrote.
6. 6
   Run the third example. If both rows are right, wire it to a surface or a workflow. If not, fix the instructions, not the access.

When the output is wrong

• Vague labels → add the label list to the instructions with one example per label.
• Made-up fields → tighten the table columns; agents follow schemas better than prose.
• Right answer, wrong format → show the exact output you want inside the instructions.
• Confidently wrong on edge cases → tell it what to do when unsure: flag, don’t guess.