alt Hacker NewsA terminal also tells the model when it failed. Exit code, stderr, a stack trace. it reads that and tries again.
The apps normal people use mostly don't. I spent two years building agents that worked inside people's gmail and calendar, and the failures that hurt all came back as 200 OK. - Slack puts {"ok": false} inside a 200. - Google ignores query params it doesn't recognise, so i sent max_results instead of maxResults for weeks and every filter was silently dropped.
The agent says "done", the user finds out later, and they stop trusting it. So i think part of the gap is that agents look a lot dumber when nothing tells them they got it wrong.
I wonder if in response to this companies like google will build a sort of agent facing layer to the app in question. So rather than you building an agent to attempt to wrangle pupeteer to click css elements as a user does, it calls some apis google exposes to certified gemini user agents and these contain exit codes and can be run as script like instructions not e.g. the shortcomings of browser agent based methods such as having to wait for all content elements to actually load and the obfuscation arms race some devs have towards exposing these elements towards headless agents.