← sources

AI agents write the code. Who debugs it when a customer hits a bug?

source: reddit (reddit:AI_Agents)published 2026-10-11 03:52ingested 2026-10-12 02:36

retained text

Right now, most of the agent hype is about writing code. copilots, coding agents & app builders.

But code that ships faster also breaks faster, and the part after shipping feels much less covered:

* A customer hits a bug and never tells you. They retry once and leave.
* The failure doesn’t throw a clean error, so there’s nothing to alert on.
* Someone still has to read logs, query the database and guess what the user clicked. For me that meant hours on a checkout bug that I never found.

So the loop looks lopsided. Writing code is getting automated, while finding out why it broke for a real user is still manual.

I’m curious how people here see it:

1. Are you using any agent for debugging or incident work, not just writing code? What does it do well?
2. Where does it fall apart: missing context, wrong root cause, or just not trusted?
3. How much would you let an agent do on its own? Read-only, open a PR, or fix and deploy?
4. Is debugging a harder problem for agents than generating code, or just a less popular one?

My take: the blocker for debugging agents isn’t the model, it’s context. The bug lives in what one customer did, not in the repo.   
  
So I’m building an AI debug agent that starts from the customer’s session, reads the code at the version that was live, finds the line and opens a PR.  
  
Do you think missing context is the real blocker, or something else?

document fetched 2026-10-12 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1x2y85d/ai_agents_write_the_code_who_debugs_it_when_a