Ask HN: How do you give estimates in the age of Agentic coding

2 points by nibbleyou 54 days ago | 2 comments

Back in the day you would get a rough estimate of how long a new feature might take once you had worked on a codebase for long enough. You knew how the internals worked, how much time it would take to design the solution, how fast you could type the code.

Now with AI doing the design and coding, it feels like there is no way to get a good estimate. It all depends on how fast the AI agent can understand your codebase, your requirements, how fast is the inference. If the LLM one-shots the best design then its fast else you keep on doing a lot of back-n-forth.

I know the estimates earlier were also not that good but at least you got a rough idea. Now at least I feel that I have no idea how long something is going to take.

It is absolutely possible that I am doing it wrong, that I am doing a lot more "vibe" coding.

Any tips or your experience will be helpful.

saltyoldman 54 days ago |

Code should be done same day, testing can take a while depending on how complex the product is. For example, some companies can simply have a unit test or integration test, but real products often need to run a 15 minute pipeline with multiple databases active and LLMs responding without errors, etc...

Our product is fairly complex, but most tickets are finished in a day.

micahdeath 54 days ago |

i ask the ai and use it - even though the code can be complete in a few minutes i spend alot of time reviewing and tweaking the output... surprising me, the timelines provided have been pretty close.