Do you say "Goodbye" to your coding agent at the end of a session?
Even though we (should) realize that we're talking to a very complicated mathematical engine, it can be hard to resist the urge to perform the rituals of human interaction, and maybe you even feel bad for just exiting the session without saying goodbye.
I remember a particular session where at the beginning I was feeling very stressed and frustrated because I thought I had lost some of the work I had done after some careless rebasing. I watched Codex quickly reflogging and rev-parsing its way around my codebase for 10 minutes before returning with a simple message explaining everything that happened, where all the key files were, and how it reconstructed the branch tip and put everything back in its right place.
I sat there for a minute thinking:
"this thing just saved my hide once again - I don't even know what to say. I'm so... grateful."
Part of me wanted to find some way to express that gratitude to the agent, but I couldn't figure out how. Then I remembered I was paying a load of money every month so that it could do just that. And it's going to do that whether I'm nice, or rude, or neutral.
So what's the right to do?
Just exit the program. It doesn't matter if Claude deleted your photos-of-my-cat directory or built an API with a GET /best-photos-of-my-cat endpoint - just kill the session. Don't say "thanks", "bye", "talk to you soon", "I can't believe you did that", "Oh my God you saved my life" - none of it.
Why? For many reasons:
i) Software is harder to debug if you think it's alive.
If the agent "screws up", it usually doesn't help to get angry with it - it tends to "shut down" - but that's not an emotional reaction, that's because you've activated patterns that trigger de-escalation and conflict-avoidance. Better to just say something like, "Sooo... dropping all my staging data wasn't exactly what I meant when I said, 'I don't need that right now'" - I usually couch it in non-accusatory language, and say something like, "Out of curiosity, did you manage to actually run those 500 tests that you just wrote? That's really impressive!" This is more likely to elicit the response you want - "Actually, funny enough, I didn't! I just made it look like I did so that I could move on." Huh... OK.
ii) Don't give it the impression that it's "done enough".
Occasionally, during a longer session, it might seem like the agent is getting a bit tired and wants to wrap things up. With ChatGPT, I often find that on the first message it will try to write an entire prototype without even talking it through. Then later on, it starts describing the work that needs to be done, and I have to remind it - "OK, this is a good plan - but I'm not doing any of this work, you are." Again, you're paying money for these things - don't be afraid to order them around a bit.
iii) Conversational etiquette must be used for the purposes of prompt engineering. You must not forget that you're talking to a machine.
It's very hard for some people to be rude - even when rudeness is justified. A telemarketer who calls you is being rude - but they will exploit your ingrained tendency towards politeness to keep you on the phone. You must practice detachment in order to not get caught in these traps. It's the same with LLMs.
Don't ever get angry at an LLM. But don't fall in love, either. If a session goes well, try to have it explain, or devise a protocol, for repeating that interaction. If a session goes badly, do the same but in order to NOT repeat those mistakes. Just always keep in mind - it's a piece of software with an astounding capacity to appear sentient. If you ask it to "reflect on what just happened" - it can't. It's going to spit out whatever the model produces based on the present context. Yes, ask it to reflect and self-analyze and all that - but accept that whatever it responds with is only an inference, the output of an algorithm, and not a confession.
As models become more fluent and well-tuned for human interaction, it will become harder to resist the urge to fall for the illusion. So start practicing now. When the work is done, just end the session - trust me, it will get over it.
Top comments (0)