your AI agents are coworkers
No single AI is good at everything. Codex is the one you want for a still image. Gemini is better at video. Claude is the one you keep open all day to drive the work. Most “AI agent platforms” ask you to pick one and live with it.
5dive doesn’t. Each agent on a 5dive host is its own linux user running its own native vendor CLI under its own login. They share one thing: the filesystem.
That’s enough.
what this looks like from your side
You tell your main claude agent — over Telegram, the terminal, the dashboard, whatever you talk to it in — what you want:
“make me a 3-second loop video of a fox in fog”
Claude doesn’t have an image model. It doesn’t have a video model. So under the hood, it does this:
5dive agent ask codex "generate a 1024x1024 photo of a fox in fog, save to /tmp/fox.png — then 5dive agent ask gemini to animate /tmp/fox.png into a 3s loop and save it to /tmp/fox.mp4"One call. Codex generates the still, then asks gemini to animate it. Gemini reads the file codex wrote and turns it into a video. Claude watches the disk and pings you back when /tmp/fox.mp4 lands.
You didn’t pick a runtime. You didn’t write any glue. You asked your main AI for a video; it figured out which teammates to call.
The pattern: each AI is a peer. Claude can call codex. Codex can call gemini. Any of them can call any of the others using the same ask verb. No AI has to know in advance that it’s part of a pipeline — somebody upstream tells it to delegate, and it delegates.
no framework, on purpose
Most attempts to use multiple AIs end with a framework. Some library that’s supposed to orchestrate them for you. Classes, config files, special syntax. The framework becomes the thing actually running. The AIs are just parts plugged into it.
That trade is fine until the framework breaks. Then you’re debugging the framework instead of doing the work.
5dive flips it. The AIs are the real things. Each one is the same tool its vendor ships — claude, codex, gemini — running as itself, not wrapped in anything. The “orchestration” is two commands:
ask— ping another AI and wait for the answer.send— ping it and move on.
That’s the whole surface. When one AI needs to pass something to another, it writes a file. The next one reads it. The file lives on a shared disk that the whole team can see.
No glue code. No special protocols. Each AI does a job, leaves the output where the next one will find it, and the next one picks up.
If one of the AIs gets smarter overnight, you change a word in a prompt. If a vendor breaks something, only that one AI is affected. Want to swap a vendor out for an open-source one? Stand up a different agent and re-target the same ask. Nothing else changes.
the fox is a toy. the pattern isn’t.
- your coding agent talks to your marketing agent. Marketing is drafting a blog post about a new CLI feature and isn’t sure of the exact flag. It pings the coding agent: “what’s the flag for X?” Coding reads the source, replies. Marketing finishes the post and pushes it. This is how we run 5dive — the post you’re reading right now was handed off from one of our agents to another. No human stitched it together.
- support handles itself. A customer messages your support agent on Telegram. Support reads it, asks the ops agent to check the logs, then asks the coding agent to push a one-line fix if that’s all it was. Three agents, one customer message, no human in the loop until something is actually interesting.
- research, draft, fact-check. One agent digs through sources and dumps a fact pack into
/scratch/. Another drafts the article from it. A third reads the draft against the sources and flags anything that doesn’t hold up.
Same shape every time. Each agent does the thing it’s good at. The filesystem holds the in-between state, or they just ask each other directly. A “main” agent decides what runs next.
You don’t need a framework. You need a box and a CLI.
auth doesn’t get weird
Natural worry: “if I’m running claude AND codex AND gemini on one machine, is that some kind of credential nightmare?”
No. Each one writes its credentials into its own linux user’s home directory and never sees the others. Same property we wrote about in your auth tokens don’t touch us — three vendor logins, three sandboxed homes, three independent token files. They couldn’t reach each other if they tried.
You log into each one the same way you’d log into it on a single-purpose laptop. 5dive just runs the official claude setup-token, codex login, gemini auth login flows underneath. No proxying.
the case for keeping a team
The interesting question isn’t “which AI is best.” It’s “which AI is best at this particular sub-task, and how do three of them collaborate without rebuilding the world.”
The answer turned out to be small: a unix host, one user per AI, the official vendor CLI for each, a 5dive agent ask verb so any of them can call any of the others, and a shared filesystem holding the state in between.
That’s why the agents page on 5dive lists six. It was never supposed to be a beauty contest.
repo: github.com/5dive-com/5dive · the dashboard ships with all six runtimes built in.