Standard Support: How AI agents ask for support

(docspring.com)

1 points | by nathan_f77 6 hours ago ago

1 comments

  • nathan_f77 6 hours ago ago

    Hi HN! I run DocSpring, and more and more of our customers use it through their AI agents. When an agent hits a bug or a confusing error, it might work around it or give up, and then we never hear about it. (I've been trying to understand why we've been getting quite a few sign-ups via Claude and ChatGPT that don't end up using the product.)

    So I wrote down a design for how an agent can ask for help, send feedback or open a support ticket, and we shipped it in our own MCP server, API and CLI.

    I often leave agents running unattended overnight. I wouldn't let them open support tickets without my approval, but suppose one gets stuck late in the evening. I might approve the ticket and then go to bed. The agent can wait for the reply in a background shell while I'm away. If the answer comes in overnight, it picks the work back up, and the work is done by the time I get back.

    The spec is CC0 and the code is MIT. I'd like to hear what you disagree with, especially on the safety side: a ticket reply is untrusted text going into someone's agent. And I'm especially concerned about the case where there might be autonomous agents on both sides of the support ticket. That feels risky. But I wonder how much longer we'll have a "human in the loop" for everything.