@emollick Interesting. What if you name the skill something else and ask Claude to call the skill as your first prompt? I have used this for other skills that do not want to load and it has worked great. It doesn’t turn it off but at least it’s a work around
@emollick Or is coding the bootstrap problem? If models become extremely good at writing software, they can build tools, evaluation systems, and automation that spill into other domains.
🤔
@cathrynlavery Better yet, have you agent be the orchestrator. So when you ask it to go
do something, it spins up a temp agent to do that thing and you can keep chatting either way your agent. Have multiple tasks at nice = multiple temp agents taking on that job. Much more secure too
@tolibear_ This is great but I will argue that it’s safer and easier to only have one actual agent you talk to (your COO). And this agent’s only job is to be an orchestrator.
One huge skill you need to give to your COO is the ability to spawn temp agents (souls and skills)
@Tesla Not if you’re on 14.2.2.3. Huge downgrade from 14.2.2.2. It’s more like hiring a 90 year old to try and drive you around…give me back 14.2.2.2 please…
@SawyerMerritt I call bullshit. 14.2.2.3 is horrible. It can’t even figure out what a turn lane vs straight lane is in this version and ignores navigation. This was a huge downgrade from 14.2.2.2
@kimmonismus I haven’t put much trust in it yet (Atlas) but use agent mode all of the time inside of ChatGPT. When you’re working on something that requires you to “do something,” agent mode works great. I also prefer that’s its sandboxed on its own system over a browser on my system.
@bobs_ter@OpenAI Actually @sama has always envisioned ChatGPT being an operating system for your life. He has mentioned this several times and you can see it’s headed in that direction.