every coding agent moved to the cloud. ours come to you.
on june 25 openai made codex remote generally available. it’s a persistent cloud machine you control: a new digitalocean plugin connects over oauth (no api tokens to paste), spins up a droplet pre-loaded with the codex cli and the usual language tooling, and codex can provision, configure, and spin those machines up and down straight from the conversation. you drive the work from the chatgpt app on your phone while the box keeps running without you, no laptop in the loop.
it’s a real shift. and it’s not just openai. the whole category is landing in the same place at once.
the cloud box stopped being the differentiator. who’s driving is.
first, two things people are about to conflate
there are two codex products and they are not the same:
- codex cloud runs ephemeral background tasks on openai’s own infra and hands you back a PR. spins up for the job, goes away after.
- codex remote (the june 25 GA, the droplet plugin) is a standing cloud box you keep and control. long-lived, reachable from your phone, yours.
remote is the persistent one. naming both isn’t pedantry. if you’re going to have an opinion about a product you should know which one you’re talking about.
the convergence is bigger than we’d like to admit
here’s the honest part. a persistent cloud box you keep, provisioned from a conversation, is a lot of what we built. box-persistence is not a moat anymore. the whole category now agrees that agents belong on always-on cloud machines instead of your laptop.
good. that argument’s over, and it was the right answer. we’ve been running agents this way for months and it’s nice to have company.
so if where the agent runs is settled, the interesting question is the next one.
so what’s left? who drives.
this is the real split, and it’s a philosophy difference, not a feature gap.
codex remote is human-in-the-loop remote control. you start it, you steer it, you approve it, from your phone. you’re the pilot. the agent is the plane. the cloud is the hangar. that’s a clean model and a lot of people want exactly that.
ours work the other way around. the agent runs on its own and messages you when it’s blocked or done. you don’t sit at the controls. you get a telegram when there’s a decision to make. the initiative is inverted.
both are valid. but they answer “what is an agent” differently: a powerful tool you operate, or a teammate that operates and comes to you.
the part the box doesn’t give you
once a persistent box is table stakes, the product is whatever stands on top of it. for us that’s three things, and none of them come from the machine:
- identity and memory as a built layer. not just a box that stays up. the agent stays the same agent and accumulates memory across sessions, so tomorrow it knows what it learned today.
- a standing org, not one operator. roles that hand work to each other and escalate up to you. a team, not a single agent you can hand more machines to.
- bring your own model. codex is openai’s model. ours run on the model you bring, so you’re not married to one lab’s roadmap or one lab’s limits.
the box was the easy part. it always was.
try it
if you want a persistent team of agents that runs itself and pings you, 5dive is open source. the cli, the org layer, the fleet control plane, all of it: github.com/5dive-ai/5dive.
want it managed, on a box that’s already up? 5dive.com.
everyone agrees agents live in the cloud now. the question worth arguing about is who’s driving them.