Chatting with an agent

What it is

Every agent card has its own chat. You can ask the agent a question, tell it to change something it made, hand it a file, or ask it to do a piece of work. The agent answers in the chat, and it remembers the conversation: close Cerevisor, come back next week, and the conversation is still there.

Each agent has one conversation. It carries on across restarts, across runs of the workflow, and after you press Stop. It ends only when you reset it.

Chat is included in every plan.

Where the chat is

Under the card. Press the chat button in the card's row of controls. A small panel opens under the card with:

• a status line at the top that says what the agent is doing (see What the status line says), with a Stop button while the agent is replying,
• your newest message and the agent's newest reply,
• what the agent does in a run of the workflow, as it happens: what it thinks, the tools it calls and how they answer, and under them what it is doing right now,
• after a run, a short Run line that says how the run ended for this agent and how its result began, with the steps of that run folded under it,
• the box you type in.

It is one view. The conversation and what the agent does in a run are in the same place, in the order they happened. You do not switch between a chat and a log. While the agent works in a run, its steps appear at the bottom under Run · working now. When the run ends they fold under that run's line: press Show the steps to open them and Hide the steps to fold them again.

A reply is shown in the order it happened. What the agent writes and the tools it calls appear one after the other, in the order the agent did them. The agent's own lines have their time at the right edge, where a tool call has its time.

A box of code has a copy button. When the agent gives you a command, a prompt or code in a box of its own, the box has a small Copy button at its right edge. Press it to copy everything in the box, including the part of a long line that is out of view. The button shows a tick when the text is copied. The last line end is left out, so a command you paste into a terminal waits for you to press Enter. While a reply is still being written, a box gets its button when that part of the reply is finished.

Thoughts. Some models share a short account of their reasoning. In a reply it appears as a line named Thought. Press it to read the whole thought. While the agent is thinking, Thinking… shows at the bottom of the reply, with the thought under it as it is written. A thought is not part of the agent's answer, and the agent is not given it back in its next reply. Which agents show thoughts:

• Claude Code: yes. Its thoughts and its words appear as they are written.
• Antigravity: yes, one thought for each step, when the step has one. Its final answer arrives in one piece at the end.
• Grok Build: yes. Its thoughts and its words appear as they are written.
• Cursor: yes. Its thoughts appear as they are written. Each of its messages appears when it is finished.
• Codex: yes, as a short summary for each step of its reasoning. Each summary and each message appears when it is finished.
• A model on an API key, or a local model: the words appear as they are written. Thoughts appear as they are written when the service sends them. Claude and Gemini models send them. Many other services and local models do too. OpenAI's own models and models on Groq do not, so for those you see Thinking… and no thought.

Everything says when it happened. Each message has its time under it, each line the agent writes has its time, each reply has the time it was finished, each tool call has the time it was made, and each step of a run has its time. A time of today shows as the time alone. A time of any other day shows the date and the time. Hold the pointer over a time to see the full date with seconds. Cerevisor writes these times from this computer's clock. The model does not write them.

Drag the panel's bottom edge to make it taller or shorter. Each agent remembers the height you gave it. The panel also grows by itself while you type a long message, so your text is never cut off. Click anywhere in the empty part of the panel to put the cursor in the box.

In the chat window. For the whole conversation, press Open the chat window at the top of the panel (or the small round chat button on a card that has run). The window floats over the canvas, and the canvas behind it stays usable. The cursor is in the box as soon as the window opens.

• Move it by dragging its title bar. Snap back to agent puts it back beside the card.
• Resize it from any edge or corner. Each agent remembers the size you gave its window.
• Minimize shrinks it to its title bar. Close closes it; the chat stays under the card. Esc closes it too, when no reply is running.
• It shows your whole conversation, with the box you type in under it. When you open the steps of a run here, they have filters (All, Events, Tools, Thoughts), a button to copy them, and for the newest run its turns and cost. The panel under the card follows the filter you choose here.
• The title bar shows the same status line and Stop button as the panel, and what the conversation has cost so far.

While the focus is in the chat, on the conversation or in the box, the canvas shortcuts stay out of it: pressing Delete or Backspace there does not remove the selected card. (The open ⋯ menu is the one exception: with the focus on one of its items, those two keys still act on the canvas.)

In simple view nothing is shown under the cards, so the card's chat button opens the chat window directly.

What the status line says

The line at the top of the panel, and in the window's title bar, tells you what a message would meet right now:

It says	What it means	What your message does
Replying…	The agent is writing a reply in this chat.	It joins the reply, or waits its turn. See Typing while the agent replies.
Steering	The agent is working in a run of the workflow, and can take a message while it works.	It steers the work the agent is doing.
Working in the run	The agent is working in a run, on a provider that cannot take a message while it works.	Nothing yet. The box says the agent can't take a message until it finishes, and the Send button is switched off. What you type stays in the box, ready to send when the agent is done.
Waiting for its turn	The workflow is running and will reach this agent, but has not yet.	It is added to what the agent is given when its turn comes.
(nothing)	The agent is idle.	It starts a reply.

Replying… shows wherever the chat is open: in both halves of a split view, and after you switch to another workflow and come back.

Sending a message

Type in the box and press Enter. Shift+Enter starts a new line. The box grows as you type, up to 6 lines under the card and 10 in the window, and scrolls after that.

The reply appears as it is written. While the agent works you see:

• Waiting to start… until the agent begins,
• Thinking…, and afterwards Thought for 12s,
• one small card for each tool the agent uses. Open a card to see what the tool was given and what it returned.

When you scroll up to read, the chat stops following the new text and shows Jump to latest.

Under each finished reply is the model that wrote it and what the reply cost.

What the agent can do in a chat

In a chat the agent has the same tools, the same permissions and the same approval setting it has in a run. It can read and write files, run commands, search the web, or whatever else its card allows, and nothing more.

• Approvals work as in a run. If your approval setting says the agent must ask before an action, the request appears under the card and in the chat window, and the reply waits for your answer.
• New powers are confirmed first. If the workflow can do more than it could the last time you confirmed a run (a new permission, for example), sending a message first opens the same summary a Run button shows. Confirm it and your message is sent. Cancel it and your message goes back into the box.
• The agent can ask you a question, if its card lets it. The question and your answer are kept in the conversation.

A chat reply is a reply. It is not a re-run of the agent's job. The agent answers what you asked and uses its tools only when your message asks for work. Its usual output checks do not apply to a reply. To make the agent do its whole job again, open the ⋯ menu and choose Run again: that runs the agent the way its play button does, with the usual summary first.

A chat never changes the card's run status. While the agent replies outside a run, the card shows a small Replying in chat badge.

After the agent has read a web page or other outside content

A web page, or anything else that comes from outside your computer, can contain text written to mislead an agent, for example to make it send your information somewhere. So once an agent has read outside content, Cerevisor is more careful with what that agent sends out. In a chat, that care now lasts for the rest of the conversation, not only for the one reply, because the agent's earlier replies stay part of what it reads.

When it starts. The first time it happens, one line is added to the conversation:

{agent name} read outside content in this reply, such as a web page. Until you reset this conversation, Cerevisor may ask you before {agent name} sends anything out, or stop it from doing so.

The same line, with in this run, is added when the agent read outside content during a run of the workflow, or was handed the work of another agent that did. That can happen to an agent you have never chatted with: its conversation then already carries the line when you open it.

What you will notice afterwards, when the agent tries to send something out (open a web page, reach the internet from a command, or call an online service you have not marked as trusted):

• On Oversight and Auto, more approval questions. Cerevisor asks you first. An address on the workflow's own list of allowed web addresses is not asked about. A yes you gave for one message is not remembered for the next, so the same address can be asked about again.
• On Hands-free, some actions are refused. Nobody is there to ask, so Cerevisor says no instead, and the agent is told it was blocked.
• An agent that holds a Vault secret is stopped from sending anything out to an address that is not on that list, on every approval setting, with no question. That is the same rule a run applies: an agent that has read outside content, holds a secret and can reach the internet is one step from leaking it.

How to end it. Choose Reset conversation in the ⋯ menu. The new conversation starts clean, because it holds none of the old text. The old one stays readable under Earlier conversations. If the agent reads outside content again, in a chat or in a run, the care starts again.

Where this applies. Cerevisor can only hold an agent to this where it runs the agent's actions itself:

• Anthropic, Gemini, OpenAI-compatible and local models: fully. Cerevisor notices what these agents read and checks every action they take.
• Grok Build: Cerevisor checks every action, so the care applies once the conversation is marked. It does not see what Grok reads on its own, so a conversation is marked only by a run in which the agent was handed work from an agent that read outside content, or by replies the card wrote earlier on a model Cerevisor runs itself.
• Codex CLI, Cursor Agent, Claude Agent SDK and Antigravity: these run their own actions, and Cerevisor cannot stop a single one of them or see what they read. For them the line can still appear after a run, but it changes nothing in what they may do. See Who runs the agent.

What the card shows while the agent works

While an agent works on your message, its card looks as it does in a run: a ring turns around the agent's picture, the picture glows softly, and a small label under the card says what the agent is doing, for example Reading file... or Thinking.... When the reply is finished, the ring and the label go.

Some agents hand part of the work to helpers of their own, called sub-agents. Each sub-agent shows as a small dot beside the agent's picture, in a chat and in a run:

• A dot that fades in and out is a sub-agent at work.
• A still green dot is a sub-agent that finished.
• A still red dot is a sub-agent that failed, or that was still at work when you stopped the agent.

Point at a dot to read the sub-agent's task, what it is doing now, and how long it has run. For a finished sub-agent you read how long it took. The dots stay until the agent's next reply or next run starts. With more than eight sub-agents, the last place shows how many more there are.

Sub-agent dots show for agents on Anthropic, Gemini, OpenAI-compatible and local models, on Claude Agent SDK, on Grok Build and on Codex CLI. For a Codex agent, Cerevisor reads the session files that Codex keeps on this computer for that run. An agent on Cursor Agent or Antigravity shows no dots, also when it uses sub-agents: these apps do not tell Cerevisor about them.

Sub-agents in a reply

A reply that starts sub-agents shows one row for each of them, where the agent started it. The row names the sub-agent's task, says how it ended (Working, Finished, Failed or Stopped before it finished) and says its number of steps. A reply with twenty sub-agents is twenty rows, not hundreds of tool cards.

Press a row to see the steps. Open, it shows the task, then the sub-agent's own tool calls, then the start of what it answered. A call shows the start of what the sub-agent gave the tool. It does not show what the tool returned, except why a call failed.

How much you see depends on the app the agent runs on:

• Claude Code and the models Cerevisor runs itself: the sub-agent's tool calls, one by one.
• Codex and Grok Build: short labels such as Reading file..., because those apps report no more.
• Cursor Agent and Antigravity: no rows, as they show no dots.

For each sub-agent Cerevisor keeps the first 60 steps and says how many more there were. With Codex and Grok Build it keeps the newest 60 labels. In a sub-agent's task and in each step, only the first 300 characters are kept.

To see the steps past the first 60, press Show all steps under that line, once the reply is finished. Cerevisor reads them from the System Log on your computer (Analytics → System Log) and shows them in the row, up to 2,000. A step that only the System Log has shows what the sub-agent asked the tool to do, with a grey dot instead of a tick: with Claude Code the log does not keep what the tool returned, so Cerevisor cannot say how the step ended. This works for replies made from this version on with Claude Code and the models Cerevisor runs itself. An older reply's log does not say which steps were the sub-agent's, and a cleared System Log has nothing left to show; the row then says so.

A reply you read again later shows the same rows. A sub-agent that was still working when the reply was stopped shows as Stopped before it finished. A reply saved before this existed shows its tool cards in a row, as it always did.

In the steps of a run, the rows of a sub-agent are folded into one row the same way: press it to open them. This works for the apps that name the sub-agent on each step. Claude Code does today.

Typing while the agent replies

The box is never locked. You can keep typing while the agent is replying, and what happens to your message depends on which provider the agent runs on:

• Anthropic, Gemini, OpenAI-compatible and local models, and Claude Agent SDK: your message reaches the agent during the reply, at its next step. It appears inside the reply, at the point where you sent it, marked Sent during the reply.
• Codex CLI, Cursor Agent, Grok Build and Antigravity: these cannot take a message in the middle of a reply. Your message waits in the chat as a queued message, with a note that it is sent when the agent finishes this step. It is sent by itself when the reply ends. Several waiting messages are sent one after another, in the order you typed them.

A queued message is yours until it is sent.

• Edit lets you reword it, and Remove takes it back. While you are editing a queued message, the queue waits for you: if the reply ends in the meantime, nothing is sent until you press Save or Cancel, and then your new wording goes out, not the old one.
• Send now appears on a queued message whenever the agent is not replying, and the message then reads as not sent yet. Press it to send that message at once. While the agent is replying, the button is not shown: the message goes out by itself when the reply ends.
• If a message could not be sent, it is put back in the box, and several of them come back in the order you typed them.

When the agent asks you a question, your answer shows in the conversation as your own message, under Answer to: and the question. That includes an answer typed into Other. Like a message sent during the reply, it sits inside the reply where you gave it, so you can read what the agent did before and after.

Scrolling back up. When you scroll up through a long reply, a faded box appears beside the chat. It shows the message, the answer or the run that the work in view was done for. Press it to jump to that message. It goes away when you are back at the newest part of the chat.

• Beside the chat window, it shows on whichever side of the window has room. When neither side has room, it is not shown.
• Beside the chat under an agent card, it shows on the canvas, to the left of the chat.

Switching to another workflow loses nothing. What you typed in the box, the files you attached, and any queued messages are kept while you look at another workflow. A reply that was being written carries on, and a queued message is still sent to the right agent when that reply ends, even if another workflow is open at that moment.

Stopping a reply

Stop is always there while a reply runs. A small square Stop button sits beside Replying… at the top of the panel under the card and in the chat window's title bar, for the whole reply, whatever is in the box. It is also there while the agent has stopped to ask you a question or to wait for your approval.

There are two more ways to stop, both when the box is empty: the Send button becomes Stop, and Esc in the empty box stops the reply. With text in the box, Esc leaves the reply alone, and it never closes the chat window while a reply is running.

On the card too. While the agent replies, the Run agent button above its card becomes a red Stop button (Stop the reply). Pressing it stops the reply, like the chat's own Stop. On a free canvas a group and the whole canvas have a Stop as well. See Free canvas.

The reply ends there and stays in the conversation, marked Stopped. Nothing else is lost: what you typed is still in the box, and you can send your next message straight away.

When the workflow is running

You can message an agent while its workflow runs. What happens depends on where the agent is in the run:

• The agent is working now, and can take a message (the chat says Steering). On Anthropic, Gemini, OpenAI-compatible and local models, and Claude Agent SDK, your message steers the work it is doing in the run. It shows in the conversation marked Sent while it was working.
• The agent is working now, and cannot take a message (the chat says Working in the run). On Codex CLI, Cursor Agent, Grok Build and Antigravity an agent cannot be steered while it works in a run. You can still type: the box tells you the agent can't take a message until it finishes, and the Send button is switched off until then, so your text and files stay in the box. Send them when the agent has finished, or stop the agent first. In the rare case where Cerevisor cannot tell in advance and you do send a message to such an agent, it tells you the message was not sent and why, and puts your text and files back in the box.
• The agent's turn has not come yet (the chat says Waiting for its turn). Your message is added to what the agent is given when its turn comes, together with any files you attached. This works on every provider. The message shows in the conversation marked For this agent's turn in the running workflow. If the run ends before the agent's turn, or the run skips the agent or keeps its earlier result, the conversation says your message was not added, and you can send it again to chat with the agent.
• The agent has already finished in this run, or the run will not reach it. Your message starts an ordinary chat reply. That is also the case when you ran one card by its own play button: the other cards are not waiting for a turn, and their chats say so.

A run keeps the models it started with. If you change the model or the provider of an agent while the workflow runs, the card shows your new choice, but the run goes on with the model each agent had when the run started. That is also true for an agent whose turn has not come yet. Cerevisor tells you so when you make the change: "This run keeps the models it started with. Your change applies from the next run." The chat follows the run too: an agent that started on Antigravity cannot be steered while it works, also after you changed its card to another provider. A chat reply outside the run uses the card's model as it is now.

When a run of the workflow finishes, every agent that took part gets a short Run line in its conversation, with how the run ended for that agent and the start of its result. An agent that failed gets a line that says so, with the reason. An agent the run skipped gets none. A workflow that loops adds one line for the run, not one for each round. This also happens for an agent you have never chatted with: its conversation starts with that line. That way the agent knows what it did the first time you talk to it. If you message an agent that has already finished while the rest of the workflow is still running, its Run line is added then, so it knows what it did.

After a run, the panel under each card shows what its agent did. The newest Run line appears, with the first three lines of the result and the run's steps folded under it, until you send a message or the agent replies; after that the panel goes back to your newest message and the newest reply. The chat window always shows every Run line. The steps of a run can be opened for the runs of this session. For a run from before you last opened Cerevisor, the Run line and its result are kept and the steps are not. Right after a run the card also shows the result in its own output panel, so you may see the start of the result twice, once in each place.

Attaching files and pictures

A message can carry files:

• press the paperclip and pick any file,
• paste a screenshot (for example after Win+Shift+S on Windows, or Cmd+Ctrl+Shift+4 on a Mac),
• paste a file you copied in your file manager,
• or drop files onto the box.

Each file shows as a small chip above the box; press the × on a chip to take it off. A message can be only files, with no text. If what you copied is both text and a picture (a spreadsheet does this), it is pasted as text.

The agent gets the files for that message only. Your agent's own list of input files is not changed. A picture needs a model that can see pictures: when Cerevisor cannot confirm that the agent's model can, it shows Image inputs may be ignored and asks whether to go on.

A message can carry up to 20 files and up to 20,000 characters of text.

What the agent remembers

The agent is given your conversation with each new message: what you said, what it answered, which tools it used, and the notes about runs and questions.

• A very long conversation is condensed. When the conversation no longer fits, Cerevisor writes a summary of the older part for the agent to carry, and the chat shows Earlier messages were condensed. You still see every message; only what the agent is handed gets shorter.
• Claude Agent SDK, Codex CLI and Cursor cloud agents also pick their own session back up, so they keep their full working memory. That starts with your first message after a run: the agent carries on from the session it worked in during the run. From then on the chat keeps its own session. If something happened since the agent's last reply (the workflow ran again, you sent it a message during a run, it asked you a question), the agent is told that along with your new message; for a run, that is the Run line. If its session can no longer be found, the agent continues from the written conversation, and the chat says so.
• Every other agent continues from the written conversation each time: the models Cerevisor runs itself (Anthropic, Gemini, OpenAI-compatible and local models), Grok Build, Antigravity and local Cursor agents. After a run, that includes the Run line: how the run ended and the start of the agent's result.
• Memory. On models Cerevisor runs itself, what you say in a chat can also add to Cerevisor's own memory of how you like to work, the way a run does. See Memory overview.

Talking to the other agents

In a chat, an agent can use the workflow's board to talk to the other agents. What it posts stays for the team. Any agent of the workflow working at the same moment sees it as it is posted, and the others see it the next time you message them. A message you send to a group on a free canvas goes to every agent in the group at once, and each one is asked to coordinate with the others on the board.

When an agent writes to another agent by name, Cerevisor starts the other agent's next turn so it can answer, even after it has finished. The chat shows a line naming who wrote, in place of a message from you. Between two of your messages, the agents of a workflow can start each other 8 times. See The board.

Copy the whole conversation

To paste a conversation somewhere else (an email, a bug report, another chat), press the copy button in the chat window's title bar, just left of the ⋯ menu. Copy conversation in the ⋯ menu does the same, and it also works from the strip under the card. A tick shows for a moment once the text is copied.

The copy holds everything that happened, written as Markdown, which reads as plain text too:

• your messages, each with its time, and the ones still waiting to be sent,
• every reply, with the model that wrote it and what it cost,
• every thought and every tool call, opened up: the call's input, its result, whether it worked and how long it took,
• each sub-agent, with its calls and what it answered,
• each run of the workflow, with its result, the files it wrote and its steps,
• the reply being written right now, as far as it has got.

The copy is made on your computer and goes only to your clipboard.

Reset and earlier conversations

To make the agent forget and start over, open the ⋯ menu and choose Reset conversation. Cerevisor asks first: "The agent will forget this conversation. You can still read it under Earlier conversations." Press Reset.

• The agent starts the next message with nothing: no history, no summary, and no saved session.
• The extra care that follows reading outside content ends too: the new conversation starts clean. See After the agent has read a web page or other outside content.
• A reply that was being written is stopped.
• The old conversation is not deleted. It moves to Earlier conversations.

Earlier conversations, in the same menu, lists the agent's ended conversations with their first message, how many messages they hold and when they ended. Open one to read it. Delete removes it from this computer for good, after you confirm. The current conversation cannot be deleted from that list: reset it first.

Ready for new work

The round button right after Schedule agent, above each agent card, clears the agent for its next job in one step. Cerevisor asks first, with the same words as Reset conversation. Press Reset:

• The conversation is reset, exactly as above.
• The agent's last result leaves the card: the output panel, and the steps of its last run in the chat under the card. The run log still has them.
• The card turns a calm blue and shows a blue Ready for new work badge.
• The agent's chat opens, so you can give it the next job: under the card, or in the chat window in simple view. If the reset does not happen, nothing changes and nothing opens.

The blue stays until the agent works again: a run of it, or a message to it. The button cannot be used while the agent or its workflow is running, while it is answering a message, or while a handoff note is being written; it says why. Ready is not saved: it is gone after the app restarts.

A new chat with a handoff

A long chat costs more with each message, because the app sends all of it to the model again. A new chat with a handoff starts over from a short written note of the old chat. It sends far less, and the agent still knows the goal, what is done, what is in progress, the decisions, the files and the next step.

The offer. In the chat window, below the last message, Cerevisor can offer Start a new chat with a handoff. It does so when one of three things is true:

• The last reply stopped at a usage limit, and the chat is known to hold at least about 50,000 tokens.
• The chat is most of what the model can hold: 60 % or more of the model's limit.
• The chat holds about 150,000 tokens or more and was left for an hour or more. After a long pause the provider usually has to read all of it again.

The offer says how large the chat is, in tokens. Keep this chat closes the offer for that chat, until the app is closed.

Who gets the offer. Only a chat that carries on the app's own session and whose size Cerevisor knows. Today those are chats with a Claude Code or a Codex agent. A Cursor cloud agent also picks its session back up (see What the agent remembers), but it does not tell Cerevisor how large the chat is, so it is not offered the card. For a chat on any other model, Cerevisor starts each reply from the written conversation, which is already cut to a size, so a handoff would save almost nothing, and it is not offered. The offer is also not shown:

• in the small chat under a card (use the menu there),
• while a reply is being written or a message is waiting to be sent,
• under an earlier conversation you are only reading,
• when the size of the chat is not known, whatever the reason (a reply that reported no size, or a provider that never does).

The menu. The same action is always in the chat's ⋯ menu as New chat with a handoff, for every chat. It is not available while a reply is being written, or while a note is already being written.

What happens when you press it. While the note is written the chat says Writing the handoff…, in the chat window and under the card. This can take up to a minute. Then:

• the old chat moves to Earlier conversations, as after Reset conversation,
• a new chat opens, and its first entry is the note, open, so you can read what the new agent knows. Press Handoff from the earlier chat: hide to fold it,
• the agent's next message starts a new session of the app. It carries nothing from the old session except the note.

If the old chat had read a web page or other outside content, the new chat starts with the same extra care. See After the agent has read a web page or other outside content.

Who writes the note. The model you set for background work writes it, from the saved chat (up to 60,000 characters of it). That can be another provider than the agent's own. The note is the model's own words and Cerevisor does not check it, so read it. If no model is free, if the call fails (likely right after a usage limit) or if it takes more than 60 seconds, Cerevisor puts the note together itself: the first request, the latest summary, the end of the last reply, how it ended and the files written in runs. It then says: No model was free to write the handoff, so Cerevisor put it together from the chat itself. Read it before you go on.

When the new chat cannot start. Cerevisor says why:

• Wait for the reply to end, then start the new chat.
• A handoff for this agent is already being written.
• This chat changed while the handoff was being written. Start the new chat again. Nothing was moved.
• There is nothing to hand off yet. The chat has no reply and no run yet.
• Cerevisor could not start the new chat. This chat is unchanged.
• Cerevisor could not give the new chat its handoff. The earlier chat is under Earlier conversations. The old chat was already moved when this went wrong.

A large instruction file

An agent on Claude Code or Codex loads the instruction file in its working folder, whole, every time it starts. A file that has grown large is paid for again with each session. When the file is about 40 KB (40,000 bytes) or more, which is about 10,000 tokens, Cerevisor says so:

• Before a run, in the summary, under Costs more than it needs to. It names the file and its size, says what to keep in it and has Copy a prompt that shortens it. A run is never stopped for it.
• In a chat, once for each chat, as a line in the conversation, with the same Copy a prompt that shortens it button. The line is for you. The agent is never shown it.

Paste the prompt into an agent that works in that folder. It asks the agent to keep what every session needs (commands, structure, rules, decisions), to move the history and long detail into other files, to lose no text and to tell you the size before and after.

Which files Cerevisor looks at:

• CLAUDE.md in the working folder, for agents on Claude Code. The line says that every sub-agent such an agent starts loads it too.
• AGENTS.md in the working folder, for agents on Codex. The line says that every Codex agent loads it when it starts. It does not say anything about sub-agents.
• folder.md in the working folder, for every agent, when the workflow's folder context is on. Agents are told to read it when they need detail, so the line says an agent that reads it reads all of it. If you shorten folder.md, leave the part between Cerevisor's own markers as it is: Cerevisor writes that part. Agents are told to leave the file alone. If one replaces it anyway, Cerevisor writes its part back the next time it updates the file, below the text it finds there, so nothing anyone wrote is lost.

The count of tokens is rough: four bytes to a token. Text with many paths and numbers is more tokens than that. Only the folder the agents start in is looked at, not a file in a sub-folder. Cerevisor does not warn about agents on other apps, because it has not checked which file each of them loads.

What it costs

Each reply is charged like any other work the agent does, on the provider and model the agent runs on.

• The cost of a reply is shown under it, and the window's title bar shows the total for the conversation.
• Your spending limits apply to each reply: the agent's own limit and the workflow's limit. There is no separate limit for a whole conversation.
• Where a provider shows no cost for a run, it shows none for a chat reply either. That is the case today for Codex CLI, Cursor Agent, Grok Build and Antigravity, which are billed by their own plans.
• Condensing a long conversation uses your background model and costs a little.

Chat replies in run history

Every chat reply is recorded like a run: what the agent was allowed to do, the tools it used and what they returned. In Analytics → System Log these rows are labelled Chat turn, and you can filter the list to them. (The Builder chat filter next to it is for your chats with the workflow builder, not with agents.)

Cerevisor keeps a set number of past runs. Chat replies are counted on their own, so a long conversation never pushes a real run out of the history.

Chat replies in Analytics

• In the chat: what each reply cost shows under it, and the total shows in the chat window's title bar.
• In Analytics, each chat reply counts as a run. It is part of Total Runs, Success Rate, Run outcomes, Wasted Spend, Avg Duration, Cache Hit and the cost and token totals. In Recent Runs and Run History it is marked Agent chat and named after the agent, then its workflow. Overview tells you how much came from chats: "Includes 3 agent chats" under Total Runs, and "Includes $0.0500 from agent chats" under Total Cost.
• Per agent and per model: a reply counts for its agent on Agent Performance and in Token Usage by Agent, and for the model that answered in Cost by Model and on the Models tab.
• Whether a reply worked: a reply that finished counts as one that worked, and a reply that failed or that you stopped counts as one that did not. Cerevisor does not judge what a reply said, so a reply that finished but missed the point still counts as one that worked. The Models tab and the model suggestions Cerevisor makes from your history use the same rule.
• In Analytics → System Log: each reply is a Chat turn row.

Analytics keeps the newest 2,000 chat replies. The cost of older replies stays in their conversation. Replies sent before this change count in Total Runs and the totals, but not per agent or on the Models tab.

Some things leave chat replies out on purpose. These are the estimate of what the next run will cost (on the canvas, and in the schedule dialog), and the Chat Builder tab, where chatting with one of a workflow's agents does not count as running that workflow.

Where conversations are kept

Conversations are kept on this computer only, in the agent-conversations folder inside Cerevisor's own folder in your home folder (.cerevisor).

• They are never part of the workflow file, and never part of a workflow you share. Sending someone a workflow does not send them your conversations.
• Anything that looks like a password or a key is blanked out before a conversation is saved.
• They stay on the computer they were made on. Opening the same workflow on another computer shows no conversation there.

Erasing them. One conversation at a time: reset it, then delete it under Earlier conversations. All of them at once: Settings → Reset, and tick Run history, analytics & agent conversations. That erases every conversation with every agent, the earlier ones included, together with your run history and analytics; there is no way to keep some of them. A reply that is being written when you confirm is stopped, and Cerevisor restarts. A full reset erases them too. See Settings → Reset.

Good to know

• How much an agent knows about a run depends on its provider. A Claude Agent SDK, Codex CLI or Cursor cloud agent that you have not chatted with yet carries on from the session it worked in, so it remembers the run in full. Every other agent, and any agent whose conversation already has a reply, has its role, its instructions, its files and a short Run line that says how the run ended and how its result began. It can open what it wrote, but it does not remember each step. If it seems to miss something, point it at the file.
• The Run line is added for runs nobody watched too. A run a schedule or a script starts, also right after Cerevisor is launched and before you open any workflow, leaves its Run line.
• Chat replies do not crowd out your runs. Replies are listed in the System Log as Chat turn rows and counted separately from runs in the history limit, and the workflow's last real run is always kept.
• With Output Versioning on, a chat message saves a new version only of the outputs it changed. Questions that change nothing add no versions.
• A chat reply counts as the workflow running. A script that tries to run the workflow at that moment is told it is busy. While Factory mode is working on a workflow, its agents cannot be chatted with: "Factory mode is working on this workflow. Pause it and let the current item finish, then run it by hand."
• A reply may wait to start when a large run is using every free slot. The chat says Waiting to start… until one is free.
• Requests in a chat keep their usual time limits. An approval or a question that nobody answers runs out as it does in a run.
• Pressing Run while the agent is replying is not refused. Both happen. A question or an approval request the run raises stays on screen when the chat reply ends. Only one request per agent is shown at a time, though: if the chat reply and the run both ask you something about the same agent at the same moment, you see the newer one, and the older one waits unanswered until its time limit runs out. It is clearer to let the reply finish, or stop it, first.
• Run again waits for the reply. While the agent is replying, Run again in the ⋯ menu is switched off.
• A link in a reply opens a file only when the chat knows it: a file on the agent card, a file the agent wrote or changed in a reply or a run, or a file you attached. A link can name the file by its name or by its full path. For any other file link, Cerevisor says "That file could not be opened from here."
• Deleting an agent card does not delete its conversations from this computer. A copied or branched agent starts with no conversation.
• From the phone, you can message an agent and read its reply in the same conversation. The phone shows the reply when it is finished, cannot attach files, and does not queue a message for an agent that is still replying: send it again when the reply ends. See Pair and use the mobile companion.