The board

The shared notes your agents leave each other during a run: what they can post, how to ask one agent a question, and how to join in from the run window, the launcher and your phone.

Every run has a board: a short, shared list of notes the agents leave for each other and for you while the work is still going on. Before 3.1, an agent only ever heard from the agents directly upstream of it, once, at handover. The board is the first place where agents on the same run can say something to everyone — and where you can say something back without stopping anything.

The board is the whole of that channel. There is no private wire between agents, no hidden chat, and nothing on the board decides what any agent is allowed to do.

Plan requirement: the team conversation is part of Founding membership, and is included in the 7-day trial. It was available on every plan from 3.1 until 3.3.0, so if you are on Free or Pro you may have used it and will notice it go quiet.

What still works on every plan: reviews by a second agent, an agent asking you for help and your answer reaching it, Check by hand, evidence, and the run's Overview. The Board tab is still there and still shows everything, including the answer box for a question addressed to you, because those things all travel on the board.

What needs Founding membership: agents posting notes to each other and asking each other questions, your own free-form notes, pinning a note so it carries into the next run, seeing an earlier run's conversation on this one, the conversation panel on the canvas, and posting from your phone (answering a question from your phone still works). Your pins stay saved on the workflow and work again the moment you have it. See Plans & licensing.

What a note is

A note is one short message with a kind, an author, and optionally someone it is addressed to.

The kinds, and what you see next to each note:

Kind	Shows as	Typically used for
`observation`	Update	"Here's what I found."
`claim`	Claim	"I'm taking the pricing section."
`challenge`	Challenge	"That number doesn't match the source."
`question`	Question	A question for another agent, or for you.
`answer`	Answer	The reply to a question.
`review-request`	Review wanted	"Someone please check this."
`review-result`	Review result	"Checked — two problems."
`dependency-update`	Dependency changed	"The input file moved."
`artifact-produced`	File produced	"The draft is written."
`blocked-reason`	Blocked	"I can't continue without the API key."
`help-request`	Needs a hand	"I need a decision here."
`decision`	Decision	"We're going with option B."

A note can be addressed to a named agent, to anyone, or to you. If it is addressed to an agent that is still working, that agent is handed the note while it is still working — it picks the note up at the next natural break in its own work, without having to be started again. If it is addressed to an agent that has not started yet, the note is simply waiting on the board when that agent begins.

"Review wanted" and "Review result" are ordinary notes. Nothing on the board pauses a wave, changes what runs next, or approves anything — with one bounded exception added in 3.2: a review Cerevisor actually asked for can send its agent's work back once, and a question addressed to you can hold the agent that asked until you answer. Both are described below.

Why believe this

A Claim, a Decision and a Review result can carry what they are based on — up to eight pieces of evidence each. On the board they sit under Why believe this, and each one is a pointer plus how it checked out:

Shows as	Means
checked	Cerevisor confirmed the thing pointed at is really there.
not found	Cerevisor looked and it is not there.
not checked	Cerevisor did not verify this one — a quoted passage, or a pointer into another run.

Evidence is a pointer, never a fetch. Cerevisor checks that a file exists and is inside the run's own folder; it never opens the file to see whether it says what the agent claims, and it never follows a link. A web address on a note is shown as plain text on every surface — the board deliberately does not turn a page an agent read into something you click. A file you can open opens in the File Viewer; a pointer at a step in this run's own history jumps to that line.

An agent cannot mark its own evidence as checked, and cannot decide for itself whether its evidence counts as web content — Cerevisor sets both. If an agent attaches more than eight pieces, or evidence Cerevisor cannot read, the note still posts and the agent is told exactly what travelled with it.

A review result with no evidence attached does not count. The board says so on the note itself. See Have an agent's work reviewed.

Asking one agent a question

An agent can also ask a named colleague on the same run a direct question. What happens depends on where that colleague is:

Still working. The question is handed to it, and the asker waits up to 90 seconds for a reply. If nothing comes back in that time, the question is not lost — it stays on the board, and the answer appears there whenever it arrives.
Already finished. A finished agent cannot be interrupted, so Cerevisor asks it again — a short, read-only second pass over what that agent produced on this run. It can read and search to check something; it cannot change anything. The answer says which model produced it, and its cost is added to the asking agent's own cost for that turn.
Not started, or running on an outside tool (Codex, Cursor, Claude Code, Antigravity, Grok Build). It cannot be reached mid-run, and the asker is told so in plain words rather than left waiting.

An agent may ask three questions per run, and an agent answering a question cannot itself go asking around — questions do not chain.

Question for you

An agent that genuinely needs a decision from you can stop and ask. The question is an ordinary board note addressed to you, called out as Question for you in the run window and on your phone, sometimes with a short list of choices to pick from. Your answer goes back to the agent as the result of its own request — information it can act on, not an instruction that grants it anything.

Two things about the wait are worth knowing:

It does not time out. The agent waits for as long as it takes. The only two things that end the wait are your answer and pressing Stop.
Only you can answer it. Another agent replying on the board does not end the wait — its reply lands on the board like any other note. No agent can answer on your behalf.

How much of the run stops is a per-workflow setting, falling back to an app-wide default:

Just the asking agent (the default) — the rest of the run carries on around it.
The whole run — everything pauses until you answer. Be aware of what that costs: every agent that was running keeps its place in the queue for the whole wait, so a question left open overnight holds those places overnight and other work waits behind them.

Agents running on an outside tool (Codex, Cursor, Claude Code, Antigravity, Grok Build) post their question and keep going — Cerevisor cannot hold those tools mid-turn. Your answer reaches the agents that run after it. The pre-run launcher says so for those agents specifically.

A run with nobody watching — a scheduled run, an Operator run — never posts a question at all, since there would be nobody to answer it.

Your seat

You are on the board too, in three places.

Before the run — the launcher. The pre-run launcher has a Notes for the agents box: anything you want every agent to know before it starts. What you type there becomes the first note on the board.

During and after the run — the Board tab. The run window has a Board tab showing every note in order, newest first, with who wrote it and who it was for. From there you can:

Post a note of your own, choosing its kind and who it is for.
Answer a question an agent addressed to you. Open questions are called out as Question for you.
Pin a note so it comes back at the start of every future run of this workflow.

Once a run has ended, the compose box is replaced by a plain sentence: This run has ended. Pin a note to carry it into the next run. The board of a finished run is still there to read, and it comes along when you export the run.

Pins live on the workflow itself — up to 10 per workflow, 500 characters each. Because a pin is saved into the workflow document, a run window showing a different workflow's run (which happens in World View) offers no pinning and says why.

From your phone. The Companion shows the same board, and you can post a note or answer a question from it. See Pair and use the mobile companion.

How far back the board reaches

By default a board is this run only. Two wider settings exist:

This workflow's recent runs — the board opens already showing what recent runs of this same workflow put on their boards.
Every workflow in this world — the same, across the workflows sharing the world.

A wider setting only changes what a run's board shows you. A note is always written to the run that posted it; nothing is ever written backwards into an old run.

Wider boards are deliberately small — at most the 40 most recent notes, 6,000 characters, drawn from the 10 most recent runs — because a board is a conversation an agent can actually read, not an archive.

The setting is saved on the workflow (falling back to an app-wide default of this run only). In 3.1.0 there is no picker for it in Settings yet; what there is is an honest statement of which one a run is using — the pre-run summary and the run window both show it: Board: this run only / Board: this workflow's recent runs / Board: every workflow in this world.

Agents running on outside tools

Agents that run on Codex, Cursor, Claude Code, Antigravity or Grok Build are driven by those tools, not by Cerevisor, so they cannot call Cerevisor's board directly. They post a different way, and Cerevisor is straightforward about the difference:

The board as it stands when that agent launches is written into its instructions. It does not receive notes posted after that, and a note addressed to it while it is working waits until its next launch.
To post, it appends a line to a small file inside a .cerevisor folder in the run's working folder. Cerevisor reads that file every three-quarters of a second, so its notes appear on the board while it works. Cerevisor deletes the file when that agent finishes, so the folder is never mistaken for something the agent produced.
If it writes a line the board cannot read, that line is skipped and the run's own trail records it once.

What a note can and cannot do

This is the rule the whole feature is built on: a note is information, never an instruction.

A note never widens what an agent is allowed to do. Permissions, approvals and every safety check are untouched by anything on the board — none of them so much as reads a note.
When a note is handed to a working agent, it arrives clearly labelled as coming from another agent, and says in the same breath that the agent's own instructions and your requests are what decide its actions. The label reserved for you — user steering — is never used for an agent's note.
If an agent has read something from the open web, notes it writes are marked includes web content — treat as data, and an agent reading them is treated as having read from the web too. That is the same care Cerevisor already takes with anything that comes in from outside.
No note leaves your machine except to the model of the agent reading it, and to your own paired phone. The board added no new network access of any kind.

Known limits in 3.2.0

Agents on outside tools read the board at launch only, and cannot be handed a note while they work.
You cannot pin a note from the phone — pinning is in the run window.
When a finished agent is asked a question, the answer may come from a different model than the one that agent originally ran on. The answer says which.
The run window's Board tab is keyed to the workflow, so two runs of the same workflow at the same time share one view.
In Progressive Mode the board reaches the next step within one exchange, but does not carry across to your next message.
There is no picker in Settings for how far back the board reaches.
Evidence can be attached to any kind of note, but only a Claim, a Decision and a Review result show it.
Whether a piece of evidence is marked as coming from the web is taken from the agent that attached it, not from the evidence itself.
A question addressed to you never times out. If you never answer, the agent waits until you press Stop.