DorkOS vs DeepSeek Harness

Each chat picks its own agent tool. Start a job on Claude Code, run the next on Codex, keep a third on OpenCode, without leaving the tab or changing any setup. All three land in one list, and they keep running for as long as DorkOS does.

This is the row where someone else matches us. Claude Code and Codex are separate plugins you install and switch on in a preset, alongside anything speaking the shared agent protocol and a second harness run as a child. Each run is one-shot: a fresh process, no carrying a session on, no way for the child to stop and ask a question, and only its final text comes back, so the reasoning and tool calls stay inside. DorkOS drives three agents as first-class citizens, with the whole session visible, steerable mid-turn and resumable. Different depth, same good instinct.

You write the job once and say when it runs: every night, every Monday, every hour. DorkOS starts the agent at that time on your own machine, then messages you when it finishes or needs a decision. Remembering is not your job.

Their documentation is admirably exact, and worth repeating in their own terms: a reminder never leaves the session that owns it, and a closed session does no work at all. Reopening one makes anything overdue arrive late. It is a way for an agent to check back on something it is already watching, not a way to have a job start at three in the morning on its own.

An agent can propose a repeating job. It never starts one on its own. You get a card naming the agent, quoting its reason, and showing the next three run times and the exact instructions. One button runs it once, supervised, before you agree to it forever.

Your agents share rooms the way people share a group chat. They see each other, answer each other, and hand a job along. You set how often they may reply, per room, so a busy channel never turns into a runaway bill.

Four dials cap the replies your agents may trade: how many in a row, how much of one conversation a single agent may take, how many per room each hour, and how many across every room each hour. Set them once, or room by room. No room can skip the dial that covers everything.

An agent that needs permission asks, and the question follows you: the browser on your phone, a Telegram or Slack message, or a notification banner on your Mac. Only the person you name gets asked, so it is never broadcast to every chat you have connected.

One bell holds every question, approval and finished run, with the things waiting on you above the things that merely happened. What you have read stays read on your other devices. A question left unanswered for the few minutes you chose rings your phone.

DorkOS does not replace a marketplace of editor add-ons built up over years. It adds the other half: package a working agent, with its instructions and its tools, and hand the whole thing to someone else in one command. Signing those tools in to outside services is the part still finding its feet.

This is the most impressive thing about the project. The list of swappable parts is generated from the code itself, so it cannot quietly drift from what is really there, and each entry says which part owns it and which parts use it. If you read the architecture before installing anything, start there. One limit: it connects out to other people’s tools, and does not offer itself to other programs as one.