@reach_vb I only know what half this means but it is useful!

@KristenForti3 Take this tweet send it to your codex and ask it to set up as needed

@reach_vb You don't even need automations anymore as the agents can now send messages to other threads

@reach_vb the markdown state tracking part is the underrated bit, i run my whole knowledge base on plain md + wikilinks and the model navigates it better than any structured db, plain text just works

@reach_vb I think it's the wrong mental model. We need solutions that don't use worktrees but cloud sandboxes.

@reach_vb Doing exact this at the moment. Inspired by symphony blog. So useful.

@reach_vb It’s my favorite feature of codex now

@reach_vb starting to feel like u need a sub-agent just to manage the other sub-agents

@reach_vb The shift from "prompt the agent" to "design loops that prompt agents" is exactly the mental model switch most devs are still missing. Worktrees for isolated feature branches is underrated too — parallel agent streams instead of sequential blocking.

@reach_vb This is the big shift people miss: the product is not just the model, it's the operating system around it. Skills + worktrees + automations compound because they preserve context across runs.

the state layer is the part that bites. simple markdown holds until the loop runs unattended for hours, crashes mid-task, and redoes work it already finished. once you want clean resume youre rebuilding real state and idempotency. thats where autonomous loops actually break not the prompting

@reach_vb also, I think this : "you can distill this all in one thread and have it spin up projects/ threads as needed" only works locally , not remote , or ?

@reach_vb Should build an abstraction to configure all of this through code. It's getting out of hand.

@reach_vb Need a token/usage cutoff step here for us mere mortals who do not have unlimited usage budgets.

I would love if you picked a rotation of team members to work within the bounds of a regular Pro sub.

I can’t work the way I want to without committing to multiple Pro subs or hoping I get hired into OpenAI.

I’m working on optimising my token use (to be open sourced soon) but really don’t think I can hit full potential without at least 2 Pro subs.

@reach_vb so different from /goal ?

@reach_vb Goat 🐐

@reach_vb Wish codex could spinup specialist subagents automatically Rn we need to specifically prompt it to spinup a subagent

@reach_vb And then you can coordinate all of this with your team via sth like http://runwork.ai which syncs these abilities between teammates and gives a deployment platform

@reach_vb At a certain point do you find that hardware is limiting the ability to run multiple loops (50+)?