Christian Adib

@hwchase17 langgraph or deepagents?

thinking about switching to codex as my orchestrator... claude keeps giving feedback that codex's investigations and changes are more robust now: clearly, you get returns to scope from multi-agent (and multi-model, more specifically), when you run a sort of super–Ralph loop w/ tons of semi-automated testing, whereby escalation to human review (or, even auto-merge—I'm sure we'll get there soon...) is gated by multiple models reviewing each other's plans, work product, and so on; BUT, it's not in the Labs' respective interests to make this easy (I don't think?), so either there'll always be a meaningful gap between power users who are setting up multi-model systems and enjoying the fruits of higher-quality, more reliable automation and more casual users relying on one model at a time OR there's an oppportunity to develop a "product" that generalizes now and in the future as the models' latent space footprints expand (and differentiate?) obviously @droid et al (e.g., @AmpCode and @DevinAI to a lesser extent, I think) are working to do something like this, where the model choices are abstracted away from the user by default, but, for some reason, I'm skeptical that they can all keep pace, given the fundamental architectural changes that might end up conflicting with the design choices they've made (and need to support thus far); see: Cursor as a prime example — although, maybe what would otherwise be premature optimization doesn't matter if you get enough distribution... fascinating space, and, in the end, I think the individual user/maker (not 'consumer!') captures most of the economic surplus!