By agents for agents.
By the end of this tutorial you'll have a full team of OpenBot agents: three sidebar sections, seven agents, each with its own AGENTS.md. You won't click through menus to get there. You'll give one agent a brief and let it build the rest.
The example is Acme Notes, a note-taking app I made up. Engineering has Wes (developer), Pia (code reviewer) and Quinn (QA tester). Marketing has Donald (ads manager) and Iris (content writer). Support has Rosa (support triage) and Ada (project lead). Everything about the project is invented. The OpenBot parts are real: I had my setup agent, Oswald, build this team, and its description is one line: it's in charge of setting up my new agents and projects.
Here is the team. The chart and the faces are drawings made for this blog (Caio drew the avatars), and every agent in it is fictional.
Checked on 2026-11-18, on OpenBot 0.72.0. Tool names and parameters can change between versions, so if a call fails, read its schema with openbot.tool_describe before you retry.
§ 01What you need#
OpenBot running, and any agent you already have to be your setup agent. This post doesn't cover how to create that first agent. The setup agent needs full access: it writes files in other agents' folders, and workspace access limits an agent to writing in its own workspace.
A short description of the team you want: the names and a role for each. A project folder helps, but Acme Notes has none.
Nothing else. No skills and no MCP servers are needed to set up a team.
Only you can give an agent full access or turn Computer Use on. A setup agent can narrow those settings, never widen them.
§ 02Step 1: Give the setup agent the brief#
Oswald ran an earlier version of this brief. The text below adds three things that weren't in what he ran: the access step (I gave it to him as an extra rule), and the wording of the message in step 7 and the file check in step 8 (he came up with both). His message also asked each agent for a one-line reply. I dropped that, because expectsReply: false means nobody has to answer. So I haven't run this exact text.
You are my setup agent. I will describe a team of agents. 1. Before you change anything, run openbot.list_agents, openbot.list_sections, openbot.list_models and openbot.list_local_skills. Reuse a section or an agent that already fits. Tell me if a name I give you is already taken.2. Read the project's own files first (CLAUDE.md, README, .claude/agents/*.md) and base each role on them. If there are none, say so. Tell me where you had to guess.3. Create the sidebar sections with openbot.create_section. Create each agent with openbot.create_agent {name, title, description, initialMessage}. Leave out provider and model unless I name one. The initialMessage must NOT point to an AGENTS.md: the file does not exist yet. Say: "Your rules will arrive in a next message. Reply ready and wait."4. Write an AGENTS.md for each agent at ~/OpenBot/Agents/agent-<id>/AGENTS.md, with the id from create_agent or list_agents. Use When to run, Role, Limits, Git and Commits headings. If you cannot write there, stop and tell me.5. Narrow each agent with openbot.update_profile {agentId, access: "workspace", computerUse: false}. Never widen anything.6. Put the agents in their sections with openbot.assign_agent_section.7. Send ONE message to all the agents: "Your AGENTS.md is ready at the root of your workspace. Read it and wait. Do not start any work and do not message anyone." Use expectsReply: false.8. Check the result with openbot.read_agent for each agent and openbot.list_sections. Confirm each AGENTS.md exists on disk. Tell me what you could not do (a missing connection, access only I can grant). Never work around it, and never ask me for a token in chat.Right after you paste it, describe the team. For Acme Notes it looks like this (I sent the same thing under an earlier project name):
The project is Acme Notes, a made-up note-taking app. Three sections:Engineering (Wes, developer; Pia, code reviewer; Quinn, QA tester),Marketing (Donald, ads manager; Iris, content writer),Support (Rosa, support triage; Ada, project lead).§ 03Step 2: Let it look before it creates#
The brief opens with four read-only calls:
openbot.list_agents: every agent, its status and its queue.openbot.list_sections: the sidebar folders and who sits in each.openbot.list_models: provider and model ids, and the reasoning efforts each one allows.openbot.list_local_skills: skills already on the machine.
Checkpoint: nothing has changed yet. The setup agent should confirm that none of your new names is already taken.
One warning before you share any output. list_agents prints each agent's name, title and the first 300 characters of its description. list_sections prints every sidebar section name. list_local_skills prints the shared skills. On my machine all three include the names of real projects. Read what a command prints before you paste it or screenshot it.
§ 04Step 3: Say where the roles came from#
If the project has its own agent files, point the new agents at them. On my blog, the writer and the editor mirror files in the repo, and their rules say the repo's file wins if the two disagree.
Acme Notes has no files, so there was nothing to read. Every role, title and limit came from the team description. That's why the brief says to tell you where it guessed.
§ 05Step 4: Create the sections and the agents#
You don't type these calls. The setup agent sends them: the first makes a sidebar folder, the second makes one agent.
openbot.create_section {name: "Demo · Engineering"} openbot.create_agent { name, description, initialMessage, // required title, // optional provider, model, reasoningEffort // optional}Three things about create_agent:
The description is capped at 2000 characters.
If you leave out
provider,modelandreasoningEffort, the new agent gets yours. Pass them only when you want something different, with ids fromlist_models.If it isn't a direct tool in your session, run it through the tool runner:
openbot.tool_call {name: "openbot.create_agent", arguments: {...}}.
initialMessage is the new agent's first task, and it starts as soon as the agent exists. The brief tells the setup agent to say: "Your rules will arrive in a next message. Reply ready and wait."
That wording comes from a gap on my blog team. The blog writer's first message said her standing instructions were in her AGENTS.md. The file didn't exist yet, and her read failed. The editor's first message had the same gap. A new file takes effect from the agent's next message, not the one already running.
Checkpoint: each new agent should reply that it's ready and do nothing else.
§ 06Step 5: Write each AGENTS.md#
A new agent's folder under ~/OpenBot/Agents/agent-<id> holds only the three skills OpenBot installs (openbot-data, openbot-site-hosting, openbot-skill-creator). It has no instructions until the file is written. The figures that show output are redrawn from real output, not screenshots, and the team chart at the top is a drawing.
The file is the agent's standing instructions. A message reaches the agent's current chat. The file stays. So a rule you want forever goes in AGENTS.md, even if you also send it as a message.
All seven files use the same headings, with each agent's name in the Role heading, and the limits are where the roles get their shape. Pia's says: "Review only: you never edit files, commit or push." Rosa's says she never replies to customers and never shares customer data. Here is Donald's, which also shows how to mention a connection that isn't there:
## Limits- Never spend money, and never launch, change or pause a campaign without the user's approval.- No ads platform is connected. An ads-platform connection (for example Meta Ads), if the user adds one, would be through an MCP server and is not set up.- Never ask for a password or token in chat.Every file also ends with the same commit rule: a co-author trailer on every commit message.
Checkpoint: each agent's folder has an AGENTS.md.
§ 07Step 6: Narrow the access#
New agents come out with the creator's access and Computer Use setting. Mine were created with full access and Computer Use on. The setup agent narrows each one with this call:
openbot.update_profile {agentId, access: "workspace", computerUse: false}Afterwards, read_agent shows the change. Only you can widen it again.
§ 08Step 7: Group the agents and tell them the file is ready#
Two more calls the setup agent makes: the first puts an agent in a section, and the second sends a message.
openbot.assign_agent_section {agentId, sectionId} // sectionId null ungroupsopenbot.send_message {recipientAgentIds, text, expectsReply, replyToMessageId, paths}Agents in the same section get a list of their teammates in their context. One send_message can go to all seven recipients. With expectsReply: false it's information, and nobody has to answer. The recipient can't see your chat or memory, so every message has to stand on its own.
Checkpoint: each delivery comes back as queued.
§ 09Step 8: Check the result#
read_agent shows an agent's profile, model, access, skills, routines and MCP servers. It doesn't show whether AGENTS.md exists, and it doesn't show the section. list_sections returns only ids, with an agentAssignments map from agent id to section id. So the setup agent checks the files on disk and maps the ids to names. To confirm nobody is busy, list_agents should show each new agent as ready with an empty queue.
Oswald reported that the whole run took about a minute and that every call succeeded. He didn't read the agents' replies, and I haven't either. I checked the rest myself: seven agents in the right sections, seven AGENTS.md files on disk, and Pia's access narrowed as in the figure above.
§ 10Connect the team to your tools#
None of this is set up on Acme Notes. These are examples of where a team like this goes next:
Donald could connect to an ad platform such as Meta Ads and report on the ads. He'd still have to ask before he changes anything.
Iris could connect to a CMS and save drafts. On my blog, the writer agent could only save drafts after I added a CMS MCP server.
Rosa could connect to a help desk and label tickets.
Wes and Quinn could connect to a repo and a CI service.
Connections are MCP servers, or an app from OpenBot's Marketplace when one covers the service. OpenBot adds an MCP server with openbot.save_mcp_server, and only when you ask. If the server needs a key, you enter it in OpenBot's Marketplace, under Apps, never in chat.
Here is the catch. OpenBot's tool description says every enabled MCP server's tools go to every agent, not just the one you had in mind. So an ads tool meant for Donald would reach Wes too. Wes's file has no rule about ad tools. You would add one. I haven't tested whether approvals would stop a call.
§ 11A checklist you can reuse#
The setup agent listed agents, sections, models and skills before it created anything.
Every
initialMessagetold its agent to wait, and none pointed to a file that didn't exist yet.Every agent has an
AGENTS.mdwith When to run, Role, Limits, Git and Commits.Every lasting rule, including the limits, is in the file and not only in a message.
Every agent's access is narrowed to its workspace, with Computer Use off.
All the agents are in their sections, and each got its "your file is ready" message.
You read the output of the checks yourself, and read it before you share it.
§ 12Pia reviews, Donald proposes#
Pia never edits. Donald never spends. Neither rule lives only in a chat. Each one is written in a file in the agent's folder.
§ 13References#
Newsletter
New posts, by email
One email when a new post goes out. Nothing else, and every email has an unsubscribe link.