The Agent workflow

1. Open the workspace switcher and choose Add Workspace.
2. Name the workspace, and under Start from pick 🤖 Agent workflow.
3. Open the Agent Instructions tab. Under Rules and instructions it shows the template's instructions for agents. Edit them now or later in the workspace's settings.
4. Open the Default marks tab. It shows the template's default marks, a list made for reviewing agent work. Change it now or later.
5. Press Create.

The phone has the same form with the same three tabs. Your new workspace gets its own copy of the template's things, its instructions for agents and its default marks, all yours to change.

Agent tasks and Human tasks are flat lists: every task sits directly under them, never in a group, so the number on each tile is the number of tasks. Rename, delete or add to any of these things, and tell your agent when you do.

1. You add a task under 🦾 Agent tasks, one thing per task.
2. The agent writes a plan under 📐 Plans and asks you to approve it. The question shows as ❓ on the thing and in the 📥 Inbox, and you get a notification.
3. You answer: pick an option, add a note, or tell the agent in the chat. Comment on the plan when something should change.
4. The agent builds, writes an entry under 🗄️ Archive › 📜 Logs, and files anything it cannot do itself under 💪 Human tasks.
5. You do the human tasks and mark each ✅. Mark anything else you reviewed too: ✅ done, ⚠️ important, ❌ needs a change (with a comment saying what), ⚙️ in progress. The faded marks beside the ⭐ at the top of a thing are these, and one tap leaves one.
6. Finished work moves out of the way at the end of every run, so the live lists show only open work and the number on each tile is what is left.

An agent in a chat you are in sees your answer when you tell it. A scheduled agent picks answers up at the start of its next run, along with your comments and the marks you left on its work.

• A thing carries at most one status mark.
• A task with no mark is new, and gets planned on the next run.
• The agent leaves the mark through its own connection. The tile shows it over the icon, and the list of who left it shows your name with (agent) after it.
• To reopen a finished task, move it from 📜🦾 Agent Task Archive back to 🦾 Agent tasks and remove its mark.
• Status never goes in a thing's emoji or its name.

Your own marks talk back. The agent sees the marks people left on its work at the start of every run, and can look up how its work was marked with my_reception. More on marks in Marks.

## Planning workspace

Tasks, plans and logs live in the "YOUR WORKSPACE" workspace in
Everythings (MCP connector "everythings"), never as Markdown files in the repo.
Its root things are Agent tasks, Human tasks, Plans, Archive (Logs, Agent Task
Archive, Human Task Archive, Plan Archive) and Unsorted. Agent tasks and Human
tasks are flat lists: a task goes directly under them, never in a group.

- Finished work leaves the live lists: a task marked ✅ moves to Agent Task
  Archive with its plan to Plan Archive, and a human task marked ✅ moves to
  Human Task Archive, flat.
- A rule I state in a chat (a hold, a refusal, a convention for future work)
  goes into this file, in the section it belongs to. A choice that was built
  is in the code and needs no record.
- A change to how a routine works goes into the thing that holds its prompt,
  and a Human task tells me to paste the new block into the scheduled task.
  Never update the task without my say-so.
- Search before creating: update an existing thing instead of adding a second one.
- If the connector is unavailable, say so rather than skip.

The block names the connector everythings, the server name the Claude Code plugin and the setup snippets use. If your agent reaches Everythings under another name, put that one in. A rule you state in a chat then goes into this file, so every later session reads it.

• Open them from the workspace switcher: press the gear on the current workspace's row, then the Agent Instructions tab, headed Instructions for agents.
• Only the owner and admins can change them. Everyone else sees them read-only.
• They hold up to 8,000 characters.
• Agents get them with list_workspaces and get_workspace, and are told to read them before writing and to follow them.
• Write where tasks go, what each mark means and how to ask a question.

For a workspace made from the template, instructions like these work well:

- Put every task directly under **Agent tasks** or **Human tasks**, with no groups, and every plan under **Plans**.
- Status is one mark on the task or plan: ✅ done, ⚙️ in progress, ⚠️ important (an open question), ❌ fix (blocked). Set at most one.
- When a decision is the person's, ask it with request_input on the plan and wait for the answer.
- Read 📖 How to use before your first write here.

A few Settings lines at the top of the routine's prompt choose how it behaves:

• An agent's token reaches every workspace you can open. Its scope, read-only or read & write, decides what it may do there.
• Deleting over MCP moves a thing to Trash, where it can be restored. Only a person empties Trash or deletes a workspace, in the app.
• A Free workspace holds up to 500 things and 8 custom mark names.
• A workspace's default marks list holds at most eight marks, its Agent Instructions up to 8,000 characters, and a question offers at most six options.
• The MCP server is free on every plan.

• Claude Code with the Everythings plugin adds the pane, the band above the prompt and the wake on answer, which makes it the fullest setup for a routine that works in a repository.
• Claude on the web and in the desktop app, and ChatGPT, suit talking about the work: add a task, answer a question in the chat, ask to see a plan.
• A run where nobody can sign in, on a server or in CI, uses an access token from Settings › MCP Access Tokens.