I feel this post. I am tired of "directing" agents when in reality it feels more like trying to herd a group of toddlers.

Sure they can mostly write better code than a toddler but this constant nudging and reminding and reiterating and stopping them from using the token budget of the whole company for a one off script. It gets tiring and I feel like I am losing brain power while doing it. Maybe it's faster but explosive diarrhea is also a faster way to produce shit.

> Maybe it's faster but explosive diarrhea is also a faster way to produce shit.

i'm in the midst of a very large global roll out of a new system including about 24 separate erp integrations. My team's system is dependent on another system coming online first then we go then the integrations come online then a massive data migration happens and ends with all the associated testing. So like a huge number of dependencies and I'm second in line. Today we get word that the first system is having issues with a vendor and will not come online as scheduled (our dry run deployment is tomorrow and spans all next week as each dependency gets their turn. Week after next is PROD.

We're not heavily integrated into that first system but we are in spots and so there's a reason why they go first. Writing around that system is feasible but we're looking around 5k lines which isn't a lot but there's still testing and sign off. My offshore team is in bed, meanwhile they, and onshore, all have other shit to do anyway so pulling them just puts everything else behind.

I tasked an agents to make the changes as soon as i woke up to the news, they got it mostly right. It took two rounds of fixes with some offshore testers who stayed up late but the PR got in and merged. We just did (literally like 5min ago) a validation run against our target system and we're green for tomorrow. I told the testers to goto bed and then i pulled up HN and opened a coke.

There's no way in hell that would have happened without the coding agents we had on tap. So, yeah, i find a lot of value in the rate at which AI can work on code even if it's not perfect every time.

AI builds like 95% for me than i play around with it for a bit for the last 5%.

I'm still a lot faster for certain things with AI but its weird. Instead of just being annoyed about some weird bug, i'm annoyed about an LLM which can do everything and then falls over itself on weird issues.

It feels weird true but it still also relevant faster?

I think the main problem is a conflict of objectives. AI companies need you to use more tokens so they are not going to improve that. They are optimizing the models to get as close as possible to the point in which people would stop using them because they are useless but without crossing that line.

Proof of this is the amount of unnecessary tasks that Claude does just as an excuse for not doing a good job doing the tasks that we ask it to do

Surely there's some point at which you can take over, and they'll have saved you time while not causing too much agony?

> herd a group of toddlers

It's called a fancy word "steering". Most of engineering is now steering or providing "taste" to AIs so they don't produce slop

I review all the agentic code and commit it myself. I also write some features manually, especially when I want to really make sure that I don't screw up. E.g. for authentication or for observability.

I also found that slowing down your development (just by a bit!) on purpose helps. Do not rush into implementing your new idea and give yourself at least a couple of days to think it through.

Welcome to management.

[flagged]