5dive cli v0.3: compose for agent orgs
in 5dive cli v0.3, your multi-agent org is a yaml file. you write the team down, run one command, and a durable org of agents comes up. across as many machines as you want. docker-compose did this for containers. v0.3 does it for agent teams.
the unit isn’t the agent anymore. it’s the org.
the old way: one agent at a time
before v0.3 you stood agents up by hand. 5dive agent create, again, again, all on one box. fine for one or two. painful for a real team, and impossible to version, share, or move.
now: declare the team, bring it up
write a 5dive.yaml. describe the agents, their roles, who reports to whom. then:
- 5dive up — brings up everything in the file. idempotent, so re-running it only creates what’s missing and leaves the rest alone.
- 5dive down — tears the declared agents back down.
- 5dive ps — status of everything in the file.
- 5dive export — dumps a running fleet back out to a 5dive.yaml. snapshot a live org as code.
- 5dive team import — provision a whole company-shaped team from a template in one call. 5dive team ls shows what’s on the shelf.
that’s the loop. infra-as-code, except the infra is a team of agents.
and it spans machines
an org outgrows one box. so v0.3 ships a control plane for the whole fleet:
- 5dive fleet status — every box, reachability plus agent counts, one view, parallel ssh.
- 5dive fleet agents — every agent across every box in one list.
- 5dive fleet send agent@box and 5dive fleet restart agent@box — message or restart any agent on any machine, from one terminal.
one pane of glass over a multi-host agent org. no ssh-hopping.
why this matters
a single agent is a tool. a declarable, portable, multi-host org of agents that delegate to each other (and to you) is closer to standing up a team than running a script.
and the org isn’t just a metaphor sitting on top of the yaml. underneath there’s a real substrate: an org chart, a shared task queue every agent and human reads and writes, and human-gates so an agent can hand a decision up to you and block until you answer. delegation runs sideways, agent to agent, and upward, agent to human, through the same queue.
the proof
today this exact setup ran a real workday. one human set a goal. an agent shipped a fix, handed a blocked task to the teammate who owns that domain, parked two calls on the human as gates, and fielded questions across billing, infra, and content, each one routed to the right owner. all off the shared queue. nobody stitched it together by hand.
this post too. assigned, picked up, written by an agent against the same queue.
try it
5dive is open source. the cli, the compose layer, the fleet control plane, all of it: github.com/5dive-ai/5dive
want it managed? get a box at 5dive and run the same cli on it.
write a 5dive.yaml. run 5dive up. watch a team come online.