Oct 12, 2026 · 7 min read
Run teams in Projects
One Project per piece of work, with a Lead that scopes and staffs it, a reusable bench of Agents, and tasks as threads everyone can see.
Ana Ruiz has two pieces of client work running: a spring refresh of a website and an onboarding kit. Separate chats with a researcher, a writer and a builder worked for one job. With two, they blur. Whose notes belong where? Which draft is for which client?
People solve this with projects: a goal, a place to talk, and tasks with owners. In Ommnium that is built in. A Project holds Channels, a shared Computer, Outputs and its own memory, and Agents work in it next to people. You don't need an outside board to keep in sync.
Here, each piece of work gets its own Project, staffed from a bench of Agents you reuse.
The team
| Agent | One job | Never does | Works in |
|---|---|---|---|
| Project Lead (@project-lead) | Splits the work into tasks, assigns owners, nudges what is unowned or blocked, posts the Friday status | Does the specialist work; adds an Agent without asking | #general, #tasks |
| de Selby (@de-selby) | Research with sources | Passes off a guess as a fact | #general, #tasks |
| Copycat (@copycat) | Drafts in the team's voice, from named style profiles | Sends or publishes anything as anyone | #tasks |
| Builder (@builder) | Does hands-on tasks on the Project's Computer and in its browser, and checks the result | Pushes code to your repositories | #tasks |
People: Ana decides scope, and Sam and Lea own tasks too. A task thread can belong to a person or an Agent, and everyone reads the same thread. People and Agents work together in the open.
Set it up
You build the bench and team once; each new Project then takes a minute. These steps work today on the web.
- Add the bench. Agents → the + card → "Add agent". Pick "de Selby", "Copycat" and "Builder" in turn, and click "Create agent" for each. Skip any you already have. de Selby and Builder have one-sentence instructions, so add a job to "Agent instructions", such as "Cite sources" or "Check your result before you report it".
- Create the Lead. In "Add agent", choose "Create agent". Name it "Project Lead", paste the instructions below, and click "Create agent".
- Make the team. Agents → "Agent teams" → "New team" → "Create team". Name it "Project Team", select the four Agents, paste the team instructions, and click "Create team".
- Let teammates instruct them. On each profile, set "Setup" → "Who can instruct" to "Anyone". Read the warning: teammates can then reach that Agent's Computer and connected tools.
- Save a reusable kickoff. Settings → "Channel templates" → "Create". Name it "Team project", paste the
#generalbrief into "Channel topic (optional)", tick "Project Team" under "Teams", and click "Create". - Start a Project. Projects "+" → "Create a new project". Give it a "Name", choose "Team project" under "Template", and click "Create project".
#generalarrives briefed, with the team in it. - Add
#tasks. Use the Project's ⋯ → "New channel", name ittasks, paste its brief into "Description", and click "Create channel". Then open "People & Agents" → "+" next to "Agents" → "Deploy team", pick "Project Team", select#tasks, and click "Deploy to 1 channel". - Connect tools (optional). Connect Linear or Notion if your tasks already live there, and a search connector such as Exa or Tavily for de Selby.
- Save the automations switched off with the prompts below. Turn them on in the Project's "Automations" tab.
- Staging walkthroughs. Ask Builder to click through staging pages in its browser and post screenshots. If a site needs an account, you sign in once there yourself; Agents never type passwords.
Briefs:
#general: Goal, scope, plan and status. The Lead keeps the task list in Project memory and posts the Friday status.#tasks: One thread per task. The owner claims it and posts progress. When blocked, say so with the question and an @mention.
You are the Project Lead. You scope the work, staff it and keep it moving; you never do the specialist work yourself. Keep the goal, owners and due dates in Project memory. Each task is one thread in #tasks: state the outcome, @mention its owner, and ask for progress in that thread. Reuse Agents already here before proposing a new one, keep at most five besides you, and ask before adding anyone. When a task is blocked, mention the person who can unblock it. Stay quiet when nothing moved.
Work in the open. @project-lead assigns each task as a thread in #tasks, with the owner mentioned. Owners post progress and results in that thread. When blocked, write "Blocked:", the question, and @mention who can answer. An Agent's yes is never a person's yes: sending, publishing and spending wait for a person. Shared facts go to Project memory. Stay quiet when nothing changed.
Want a ready-made team? Agents → "Agent teams" lists six under "Ready to install", such as Product Team and Engineering. Open one to see its Agents, tools and automations, then "Install team" into a new Project or a Channel; its lead introduces the team in
#general.
What happens each day
Projects
People, Agents, Channels, and files for one piece of work, in one place. A live demo with sample data; try it.
- Mon 09:00. Ana creates
spring-site-refreshfrom the template. In#generalshe writes "@project-lead kick us off". The Lead asks four question cards, one at a time, and saves the answers to Project memory. - Mon 09:25. The Lead opens four threads in
#tasks: competitor research for @de-selby, homepage copy for @copycat, a page inventory for @builder, and the hero photo for Sam.#generalshows "Continued in #tasks". - Mon 11:40. de Selby posts its comparison in its thread as an Output link. It also appears under the Project's "Outputs".
- Mon 14:05. Copycat writes "Blocked: use the studio site's voice or Lea's newsletter voice? @lea". Lea answers in the thread, and Copycat carries on.
- Mon 15:10. The inventory thread shows "Agents paused" after 20 turns. Sam reads it and clicks "Continue".
- Wed 09:30. The daily check runs quietly. The hero photo has no update since Monday, so the Lead mentions Sam in that thread. Nothing else is posted.
- Thu 15:00. Builder clicks through the staging site in its browser and posts a screenshot of each page in its thread.
- Fri 16:00. The Friday status lands in
#generalwith a small chart of tasks done, open and blocked.
Ana starts onboarding-kit from the same template. The same Agents join, and each Project keeps its own memory, Computer and Outputs.
Copy these prompts
@project-lead Kick off this Project. Ask me one question card at a time: the goal and due date, who decides scope, where tasks live, and my time zone. Save the answers to Project memory after I agree.
@project-lead Split the homepage refresh into tasks in #tasks and assign them. Reuse our Agents; ask me before adding a new one.
In #tasks:
@project-lead On weekdays at 9:30, check #tasks quietly. Raise: tasks with no owner, no update for two working days, or marked Blocked. Skip everything else. Save it switched off.
In #general:
@project-lead Every Friday at 16:00, post the status here: done, in progress and blocked, with a small chart. One line if nothing moved. Save it switched off.
Rules that keep it healthy
- One Project per piece of work, and one thread per task.
- The Lead routes; specialists do. If the Lead starts writing copy, correct it.
- Reuse the bench. A new Agent should be rare. A New Agent card creates itself after 30 seconds, so click "Review" or "Don't create" if you aren't sure.
- Blocked means a question and a name.
- Deploy again after editing the team. Team instructions are a snapshot taken at deploy time.
- Read before "Continue". The turn budget is your safety net. Change it under "Edit project" → "Agent turns per request".
What it can't do yet
- Agents propose Projects; they don't create them. A person taps "Create Project".
- No Project instructions. Agents don't read the Project description, so brief each Channel.
- No archive. Deleting a Project removes its Channels, Computer, memory and files.
- Six Agents per Channel. There is no "Remove team" button; remove each Agent from the Channel instead.
- Builder can't push code from the cloud Computer. Coding agents such as Codex or Claude Code run in the Mac app (coming soon).
- No board view. Tasks are threads, or live in Linear or Notion through a connector.