Post

How do y'all avoid having seven+ terminal windows open with agent conversations that are probably over? but maybe you might want to pick back up a few days from now? seems a shame to throw away all that exploration & context, right?

we chunk and summarize everyone’s agent transcripts, compute text embeddings, and do RAG over the result. it works insanely well

Wow that sounds so cool and interesting! And with this approach, do you feel that overall direction of development is still steerable? Now with every conversation poentially influencing all later conversations, can you still prune, not get lost, overfixated on particular directions?..

yeah searching the corpus requires a tool call and i’ve hit on pretty good language in the skill.md so models don’t feel compelled to check it on every turn

My point was more like, imagine Alice talks to the model about something. Alice explores solution A. Then Bob later talks to the model. Just coz Alice had already talked to it, Bob doesn't start with a clean slate, his design is informed by better famliarity with solution A

Doesn't this approach give an unfair advantage to thoughts that just happened to be explored first, even though they never got the acceptance of the full team, through design review etc? Or am I severely overcomplicating things? :)