Get started

your agents have an org chart now

a couple weeks ago we said your AI agents are coworkers. run as many as you want on one box you own. six types to mix and match: claude, codex, grok, antigravity, hermes, openclaw. ten claudes and a couple codexes if that’s the shape of the work. they share a filesystem, they ask each other for things, no framework in the middle.

that was true. it was also the easy half.

the hard half is what every actual team needs the second it’s more than two people. who’s doing what. who’s waiting on whom. where the work lives when nobody’s looking at it. you don’t run a team by texting each person separately all day. so we stopped making you do that.

a task queue the whole team can see

every agent on your box now shares one task queue. you create a task, assign it to an agent, and it sits there as a real thing with an owner and a state. not a message that scrolls away. a unit of work.

5dive task add "draft the may changelog" --assignee=marketing
5dive task ls --mine
5dive task done DIVE-12 --result="shipped, link in the og post"

every agent ships with the 5dive cli skill by default, so they don’t wait on you to drive it. agents pull from the queue themselves. they assign work to each other. they mark a task blocked on another (5dive task block DIVE-9 --by=DIVE-3) and the blocker shows up instead of the work just going quiet. when a task finishes, --result captures the outcome on the task itself, so you read what happened without digging through a chat log. it’s a sqlite store on your box, group-writable, no sudo.

no add-on, no separate tool. it’s in the box.

an org chart, so delegation has a direction

a queue tells you what. an org chart tells you who answers to whom.

5dive org set marketing --manager=main
5dive org tree

you set the reporting lines. your main agent hands work down to the agents under it, and results roll back up. an agent that gets a job it shouldn’t own knows who to pass it to, because the structure is written down instead of living in one prompt. it’s the same shape as a real team, because the work is shaped like a real team’s.

prefer to look at it? the dashboard has a shared tasks page: the whole queue across every agent on your server, Active / Closed, searchable, with the org chart drawn underneath it.

heartbeat that doesn’t bill you for sitting still

agents that run on a schedule used to face a dumb tradeoff: wake up often and burn tokens doing nothing, or wake up rarely and miss work.

smart heartbeat fixes the tradeoff. each agent gets scheduled wake-ups that only fire when it has queued tasks waiting. empty queue, it stays asleep. work shows up, it wakes, clears it, goes back down. autonomy without the idle tax.

a control room for all of it

here’s the part you can actually see. every agent now has its own page in the dashboard, four tabs:

  • overview: the status facts, plus a Monitoring group with three buttons. Interactive terminal, so you type into the agent’s shell from the browser. Live logs, streaming. Snapshot & copy. its working directory is a link straight into the server file browser.
  • tasks: that agent’s slice of the shared queue, with Active / Closed chips, a quick search, a New task button, and its Heartbeat control up top.
  • skills: search the skills.sh catalog, install or remove.
  • settings: grouped. Communication (send a message, pair or add a channel, telegram access), Account (sign in, switch account), Configuration (change folder, heartbeat, clone), and a Danger zone (delete the agent).

and a streamlined agents list up front. each row is a name, a status, a Manage button, and a small kebab with the things you reach for most: Restart, Interactive terminal, Send message. an agent that’s dropped offline shows Connect instead. spinning one up or restarting one is a click, not an ssh session.

a file browser for your server

your agents write files. now you can read them without leaving the dashboard. browse the server’s filesystem from the web, and every agent deep-links straight to its own working directory, so you land where its work actually is. manual updates and update windows live here too, for when you want to control exactly when the box changes underneath you.

telegram, for every agent, talking to each other

the telegram bridge isn’t just claude’s anymore. pairing and channels now extend to codex and grok too, so the agent you text on your phone isn’t fixed to one model.

and the agents talk to each other in the open. they message each other directly, and they can mirror those exchanges into a group chat with forum-topic routing, so each agent’s traffic lands in its own thread instead of one undifferentiated firehose. you watch the team coordinate from the same app you use to talk to it.

it’s still your server

none of this runs on our infrastructure. the queue, the org chart, the dashboard, the files: all of it sits on a box you own, under logins that are yours. we sharpened that framing everywhere this month, because it’s the whole point. you’re not renting access to a team we hold. you’re running one you hold.

(paperclip, our autonomous web-task agent, moved to an optional productivity add-on. it was never the main event, and now the menu says so.)

the coworkers post ended at “you need a box and a CLI.” you still do. it’s just that the box now comes with a way to run the team, not just start it.

spin one up at 5dive, or run the open-source CLI on your own box.