The recent vid's are great. Are y'all actually building a software factory as well for public use, or just for internal use? I am wanting to formalize more my own software factory and thinking about doing it directly in BAML, but certainly starting with some point-solutions in BAML (such as a clean-room PLAN review workflow). Just curious if there are particular lanes that would be cleaner to stay in w/r/t BAML project roadmap...

we're currently building an upgraded one! @dhil and @Antonio Sarosi (BAML 🐑) are working on it, but its not live yet! we'll likely be open sourcing it so adapting it to your usecase should be easily done

i've been thinking about this a lot; I have grown eng teams from solo to 15 eng's a few times in my career, and frankly I think software factories are more an "organizational design" problem as much as they are a logic problem. Most of good engineering delivery is due to good culture, good processes, and good org design. The models are getting better and better at figuring shit out; the more i use them the more I think they just need a few basic guardrails. The one I want to start with is hardening PLANs for big changes, by building multi-agent red-teaming. The problem I find w/current LLM is that they cannot do more than a handful of thigns at once, but they can do any individual or small set very well. So my red teams produce like 30-50 "concerns' and the model will do like 4 of them. So I want to make a lightweight internal process so that these "lists of concerns" get decomposed into a mini todo list and then the main agent can just work through them all. I think this will result it much better adherence.

one of the tricky parts of red-teaming is that ai agents will always find something wrong with some code if you ask them to find that

they overfitted on Jobs' "one more thing" for sure! i instructed it do it a bunch of times until the additional findings are marginal/immaterial and it has successfully exited the loop...

Some agentic tools and/or models are better than others at using sub-agents for what you describe, while some don't support it at all.

I like your description about "organization" design. The environment/how-to-do-things plays a lot into whether something succeeds or fails. Are you playing with agent skills or plugins to make the changes in process easier to handle?

haven't decided yet. probably plugins so that I can better coordiante things amongst agents. I kinda like the idea of installing "brain taps" on agents to look at all convo, and then proxy useful nuggets to other agents. imagine if you had someone that could read your mind and then ciruculate aporpos thoughts to others working on the same problem

I'd love to pick your brain on this at some time, if you are open. I think ability to iterate quickly would be key. Every team works differently and may evolve diverge in different directions.

hadn't heard of the DeepSeek harness... I'd be happy to chat whenever! i am pretty free these days, lmk when works