I'm on the fence about this one, @gregorriegler. I literally did this today, but it feels like it might be too inconsequential to be a pattern. @devill what do you think about this one?

Also not entirely sure the agent doesn't know it can actually execute the command itself, so don't quite agree with the framing. If you ask it directly, I think it knows and will tell you that it can.

It feels like it depends on the model a lot and also sometimes random. Last May the OpenAI models tended to do this a lot compared to Claude, tell you how to do things vs doing them. It felt more like a model training difference / less agentic training or a model less incentivized to do things on behalf of the user.

Hmmm... I do that all the time, but I never thought of it as a pattern. The difficult part with this, is that we are fairly experienced, so many things may seem obvious to us that really aren't to most people. Indeed I did have to remind the team I've been coaching recently several times that this is an option. So I do see a good reason to include the pattern.

OTOH I think we should find some time during the Samman open space in Vienna and discuss how to move forward with the patterns. We need to organise them at this point, potentially deduplicate and we should define some clear acceptance criteria, so we don't end up making these calls on a whim.

Fair enough! Good points