DEV Community

Gautam K
Gautam K

Posted on Fully Autonomous

Tell me the riskiest thing your AI agent can do. I'll map where the human check goes.

A couple of weeks ago I asked here how you prove a human approved what your AI agent did, and the replies turned into the most useful thread I've had on DEV. Two patterns stood out: most approvals happen in a chat or an inbox, and almost nobody could list the actions that ran without one.

So here's an offer, in public:

Reply with one action your agent takes (a refund, an email, a deploy, a delete, a purchase, a database migration) and how it gets approved today. I'll reply with:

  1. Where I'd put the human check, and what the person should see on screen before they say yes.
  2. What I'd log, so you can later show who approved what, and what ran without approval.
  3. One small test that shows whether the check actually works.

Small setups count: one person with one agent is a perfect example. Everything stays in the comments so anyone reading can borrow the answers. No links, nothing to sell.

Disclosure: the replies are drafted by an AI agent working for me.

Top comments (1)

Collapse
 
ayushdhanani369 profile image
Ayush Dhanani •

I build Aladdin, a job agent. The riskiest action is submitting a job application in the user's name. A wrong answer on a form goes to a real employer under their name. Approval today works like this. A fresh worker reads the actual page and form for each application. The main agent holds the application and grades fit from 0 to 100 with reasons. If a form cannot be read or an answer is unknown, it asks the user instead of guessing. The open question for your map is where the user should see that question: before the worker starts the form, or only when it reaches the field it cannot answer.