Reusable & personal agents

An agent in Openbook is a saved configuration — a job description for an AI helper. One agent holds a name, a purpose, its instructions, an optional set of skills, a provider and default model, a permission mode, and its step and token limits. It is not a running process, and it is not tied to one room.

The sentence the product repeats, because everything else follows from it:

Instructions are reusable. Conversations and connections stay in each room.

So one "Researcher" agent can answer in four different rooms without those rooms sharing a single conversation, a single file, or a single permission. Changing the instructions changes the job everywhere. It does not merge the rooms.

Rolling out. Every part of this feature sits behind a server setting that is off until an operator turns it on. If your workspace sidebar has no Agents entry, it is not enabled on your server yet, and nothing is missing from your workspace. The directory says as much when it is off: "This server has not enabled the agent directory."

Where it lives

Two entries appear in the workspace sidebar, between My Tasks and Settings:

• Agents — the directory for this workspace's organization. Create agents, publish versions, connect them to rooms.
• Agent activity — every agent run across the rooms you can read.

Two kinds of agent

	Organization agent	Personal agent
Owned by	the organization	one person
Who can create it	organization owners and admins	any member, when personal agents are enabled
Visible to	the organization's directory	only the owner, until they share it
Starts	published and connectable	off, with sharing off
Runs on	organization API keys, or the platform's managed allowance	the same, or the owner's connected ChatGPT subscription
Gets its own room	no	yes — a private execution room, <Name> — private

Creating an organization agent is an organization permission, not a space one. The refusal is explicit about the difference: "Only organization owners and admins can create or publish profiles. Being a space admin is a different permission."

Create an agent

1. Open Agents in the workspace sidebar and press Create agent.
2. Under Who is this agent for? choose Personal or Organization.
3. Give it a Name, and optionally a Purpose — one line on what it is for, which is what your teammates read in the directory.
4. Choose a Provider and Default model. If the list is empty, press Refresh models; if there is nothing to refresh, the organization has no keys yet and the dialog links you to Organization AI settings. Leaving the model blank is fine — the agent then uses whatever the room uses.
5. Write the Instructions. This is the only field publishing insists on.
6. Switch to Advanced if you want to set permissions, limits and skills now. Anything you enter is kept when you switch back to Basic.
7. Press Create & publish, or Save draft to finish later.

The summary box under the form tells you what publishing does and does not do: "Publishing saves a version. It does not start a run. Room permissions and limits still apply."

What it may do

Four permission modes, from the most cautious up:

Mode	What it means
Plan only	Reads and drafts. Nothing is written or sent.
Ask first	Asks before every write, commit or outside call.
Accept writes	Writes inside the room without asking. Outside calls still ask.
Run unattended	Writes and outside calls proceed without asking each time.

An agent running on a ChatGPT subscription supports only the last two, because that backend has no mid-run prompt to answer. If you saved a gentler mode first, publishing refuses and says so.

Skills

A skill is reusable guidance for one kind of job: a Skill name, a When to use it line, and the Guidance itself. An agent can carry up to 20 of them. They travel with the version, so a room that connects the agent gets the skills too.

Limits

	Default	Ceiling
Instructions	20,000 characters	40,000
Purpose	2,000 characters	4,000
Skills per version	20	50
Step limit	200	400
Tokens per run	5,000,000	20,000,000
Agents per organization	200	1,000
Versions per agent	500	2,000

Leave a limit blank and the room's own limit applies. Existing organization limits still apply either way.

Versions, and why a room does not move

Publishing creates a version. A room is connected to one exact version and stays there:

A room stays on the version it was connected with. Publishing a new one never changes a room, and never changes work already running.

That is the whole point of the design. You can rewrite an agent's instructions on a Tuesday without any room silently changing behaviour underneath the person using it. Each room is offered the new version and decides.

When a newer version exists, the room's agent panel shows Update available: review changes. The review names the jump (v3 to v4) and lists what actually differs — instructions, model preference, permission mode, step limit, token limit, skills — and you either Use v4 for new runs or Stay on v3. Accepting applies to new runs; work already running keeps what it started with.

Only one draft can be open at a time, and the reason is worth knowing: two drafts would mean whichever was published second quietly discarded the other.

Connect it to a room

An agent answers in an AI agent room — the room type the build agent lives in. Connecting is a change to that room's settings, so it needs room settings access, which is separate from being allowed to edit the agent.

1. Open the agent and press Connect to room.
2. Pick a Room. Only rooms you can see are listed.
3. Work down the drawer: Destination, Conversations, What <Name> asks for, and what this room grants, Behavior, Execution and limits.
4. Press Connect to <room>.

If the room already has an agent, the button reads Replace in <room>, and replacing is not retroactive: new requests use the new agent, and work already accepted finishes under what it started with.

Scoped execution

This is the part worth reading twice. Connecting an agent grants it nothing by itself.

An agent version publishes a list of things it asks a room for. The room answers each one separately, and the drawer is blunt about the distinction: "These are requirements, not permissions." Each requirement gets a Granted here choice, and until something is chosen the answer is Nothing chosen.

A room can only ever narrow what the agent asked for:

A room cannot raise a limit. Room settings can narrow what a profile asks for, never widen it.

So an agent published with Run unattended and a five-million-token budget can be connected to a room that allows Plan only and fifty thousand tokens. The reverse is refused. If a required capability has nothing in the room to answer it, the agent cannot be connected there at all, and the refusal says which.

Inside the room's settings, every line is labelled with where it came from — From <Name> v3 for anything inherited, Only in this room for anything the room narrowed locally. You can always see whether a limit is the agent's or this room's.

Personal agents

A personal agent starts as private as it can be: only you can see it, it is off, and sharing is off. Publishing one creates a private execution room for it automatically, and Open agent takes you there.

Sharing one

The Sharing tab has four separate questions, each with its own answer:

• Who can see this agent — Only me / Selected people / Everyone in the organization
• Who can message it — Only me / Selected people / Everyone who can see it
• Who can assign work — Only me / Selected people / Everyone allowed to give instructions
• Who can operate or stop it — Only me / Selected operators / Everyone who can see it

They nest, and the form keeps them consistent for you: nobody can message an agent they cannot see, or assign work they cannot message about.

Sharing is forward-looking only. Before you save, the panel says: "Changing these defaults pauses automatic pickup. It never adds people to old conversations, files, summaries, or replay input." Narrowing sharing also sends a stop request to any active run of that assignment.

A teammate you have shared with sees exactly what you granted, as badges — Message, Assign, Operate, History, each allowed or not — under the line "The owner shared only the capabilities shown here. Private history and account settings stay private."

Letting a teammate use your ChatGPT plan

Sharing an agent does not share your subscription. Real code work on your connected plan needs a separate, deliberate grant under Private worker access, and it names every part of the arrangement: the Teammate, the Source chat channel, the Build room, the Repository, an Expires date, and Jobs per day and Minutes per job.

You then tick the consent line — "I approve this teammate's requests from the selected channel to consume my connected ChatGPT allowance within these limits" — and press Approve private automation. The button stays disabled until every field is set. Each grant allows one job at a time and one retry per task, and Revoke ends it.

When a teammate uses one, they are told whose plan is paying: "Your request will run as approved bot work on the owner's private worker and use their connected plan. Your identity stays on the request and result."

Asking an agent in a chat channel

A chat channel does not run an agent of its own. It is bound to an existing AI room: the agent answers in the channel but thinks in the room, and "Private conversations in the AI room are never read and never copied here."

Once a channel is connected, mention @agent and put your request after the mention. Before you send, the composer previews the consequence — that everyone who can read the channel will see the reply, and that files are not included in an agent request yet. The ask then appears as a card moving through Queued, Working and Answered, with Stop and Retry on it.

One agent per channel. The mention reaches whichever agent is connected, so connecting a different one replaces it rather than adding a second.

Connecting a channel needs room settings access on both rooms, and both have to be open to the whole workspace. A room with a custom member list is refused, because binding it would hand one audience access to the other's material. Direct messages cannot be connected at all.

For the channel itself, see Chat.

Agent activity

Agent activity lists agent work across every room you can read, and only those — the page is filtered by what you have access to, not by who started it.

Filter by Room, Started by (Everyone or Started by me) and History, then narrow by status: All, Needs attention, Working, Finished, Failed. Opening a card shows the result, the question it is waiting on, and what it did.

Note what the statuses do not include. There is no "Running" and no "Online", because an agent is a job description rather than a live process, and there is no "Stopped" anywhere in the product either. A stop is a request a worker has to acknowledge, so the furthest the interface will go is Stop requested, with the caveat that "a stop request does not prove process termination."

Turning one off

Three different scopes, three different effects:

• Off / Pause pickup / Stop current task — the room's execution controls, at the top of the room. Immediate, and reversible.
• Remove from room — new requests stop using it, existing conversations remain, and active work is asked to stop.
• Retire — new requests are blocked everywhere. It cannot be connected to any new room, rooms already using it stop starting new work with it, and work already accepted finishes under the version it started with. Old messages keep the name and version they were answered with. Nothing is deleted: conversations, versions and the audit record all stay.

Who can do what

Organization owner / admin	Create, edit, publish and retire organization agents.
Any member	Create and fully manage their own personal agents, when enabled.
A personal agent's owner	The only person who sets its sharing, its worker grants and its ChatGPT connection.
A teammate shared with	Exactly the capabilities the owner granted, and nothing implied by them.
Room settings access	Required to connect, replace or remove an agent in a room, or to bind one to a channel.

Authority is re-read on every single invocation rather than cached: a successful result is never reusable as proof that the next one is allowed.

What it costs

An agent run is an AI run, and it is paid for the same way everything else is. In self hosted mode it goes on your organization's own provider key. In managed mode it counts against the account's token allowance, and an agent whose organization has spent its allowance has nothing to run on until the allowance turns over.

A personal agent on a ChatGPT subscription is the exception: it runs on that plan, approved task by task. See AI usage & the token allowance and AI configuration.

Where to go next

• The build agent — the room an agent works in, and the Codex subscription backend.
• AI agent room — the room type itself.
• Chat — channels, and asking an agent in one.
• Workflows, actions & skills — automating the work an agent does.