Viduck
viduck.bsky.social
Viduck
@viduck.bsky.social
Ducks and data science, not related to DuckDB.
5 - Agents are like a senior developer (that makes dumb mistakes sometimes) that has decided to read your code for 10 minutes and help you out.

As smart as it can be, it doesn’t know your domain and will need to read it all. And the more context an LLM has, the lesse effective it becomes.
June 18, 2025 at 9:35 PM
4 - Of course learning to speak their language and correctly prompt them in a step by step fashion is important.

But your stack not being too different and complex from the common training data out there is valuable and saves time. Which is the whole point of AI Agent development, to save time.
June 18, 2025 at 9:34 PM
3 - The allure of NodeJs, Python and Golang isn’t just that LLMs know it, but that they get further with less confusion and complexity.
June 18, 2025 at 9:34 PM
2 - After using .Net for 15 years (since XNA) and Java I love it as much as the next guy. But Agents that consume 2mil tokens just understanding your Mediator and DDD pattern is a receptive for failure.
June 18, 2025 at 9:33 PM
3/3 We’ll need to train everyone on which model to use, how to instruct them step by step, and how to rely on ADRs.
June 18, 2025 at 9:26 PM
2/3 Working in a data team it’s easy to forget others struggles. I keep seeing FE or .Net devs simply install Windsurf with the default (dumb) model and just give it a big task.
June 18, 2025 at 9:26 PM