I appreciate your blog posts on the topic here and hope you keep sharing them!

Related: we copy/pasted our interview take-home/in-person exercises into Claude and got great results. Rather than say "don't use it," (why? we let them use it at work!) we changed the follow-up questions we ask if someone does.

"So...how do you know it works? How would you test it?"

Here's them saying "not unless we specifically indicate you can" www.anthropic.com/candidate-ai...

(apologies if you were asking this as a more specific follow-on to this document)

This is...sort of brilliant.

The past few months, I've been Claude Coding cute lil' games for friends' birthdays ("You like triathlons and we think horses are funny, here's a side-scroller with both that ends in 'Happy Birthday!'"). Previously I'd text "Happy Birthday! I hope it's a great one!" Not art, but how is it perceived?

If you end up writing something on what you learned, I'd love to read it!

(I wish I had something to contribute, but we mostly encourage people to take the company time/money to set up and use the tools, and share success stories where they have them. Though very possible I'm dumb/out of touch :D.)

Makes sense! Thank you for staying curious and sharing with the rest of us! :)

Great read! Beyond the async vs sync distinction, is there a reason you are dabbling with so many tools? For example, are you finding Claude Code to be materially different from Codex CLI, or is it more a matter of trying them all out of curiosity? Thank you for sharing!

Great summary, thank you for sharing! I'm similarly scratching my head at "why not iterate on the agents?"

If I understood correctly, they assume the DB can leave hints for the exploratory agents, so they are already assuming more than a black box between client and server.