3 bugs I've found (so far) with the new composer

General
1

I really like the new agent functionality, but I noticed 2 bugs right off the bat:

  1. The agent often has to rewrite commands, showing an error Command contains newline characters. I work on agents, and this strikes me as something that can be easily fixed with a small prompt tweak, ideally the agent should never make this mistake
  2. I noticed that more than 50% of the time, the terminal interface that opens when an agent runs a command fails to render anything. IE: the black box opens where the command is running, and expands in size as though the command is printing output, but the box remains black and empty. Switching tabs in the editor and returning to the composer tab will cause the box to render, but this is pretty annoying.

UPDATE: I found a third error

  1. When you have an open composer tab, you switch to a different tab and switch back to the composer, it will have switched to a totally different composer than the one you had open (ie: it seems to pick one from the composer history, not clear if it’s random or not), but you will then have to open the composer history tab, select the latest one. This seems to happen every single time.

Overall the new agent interface is majorly additive, and for me I’d prefer this as the default over the base composer. I’d love to be able to add my own tools as well (but understandably that might not be a priority for you).

In fact I can barely tell the different between the base composer and the chat anymore, it seems like these could be merged.

I’d lastly like to echo the main complaint I think others have with the new composer:

  • the best part of the old composer was being able to see all the diffs that were applied very quickly. This is much harder in the new interface, however the new interface is still quite usable, it just feels like it makes the chat interface mostly redundant now.

Keep up the great work guys! Excited to see where the agent composer goes!
One bonus idea:

  • We found giving agents a general thought_pad or thinking tool that they can use to plan actions before executing them is a cheap way to provide a little chain of thought. Do with that what you will