Free canvas

What it is

Every workflow is in one of two layouts, and you can switch between them at any time:

• Waves (the usual layout). Agents sit in columns from left to right, and Run works through the waves in order, each wave handing its work to the next. This is a structured workflow, also called an ordered or deterministic workflow: the same order every run.
• Free canvas. The waves are gone. You put each agent wherever you want, gather agents into named groups, and talk to them through their chats: one agent at a time, or a whole group with one message. This is a freeform canvas: there is no order, and the workflow is not run as a whole.

Use a free canvas when you keep many agents around and work with them by talking to them, not by running the workflow from start to finish. Use waves when you want the agents to run one after another in a set order.

Turn it on

In the Workflow bar at the top right of the canvas, press the Free canvas button (three small cards set down out of line). It sits right after the Software factory button and lights up while it is on. Its tooltip reads Free canvas: place agents anywhere, group them and message them. Running the whole workflow is off while it is on. On a narrow canvas it is the Free canvas row in the More workflow actions menu (the three dots). See The Workflow bar.

When it is on:

• the wave headers disappear and every agent stays where it is,
• the lines between agents are hidden,
• Run (Play), its options, Loop and Advanced Loop are turned off, and Endless mode and Loop & Enhance refuse with a short note. A run that was already going can still be stopped.
• the workflow's frame says N agents | free canvas, and the status bar shows Free canvas · N groups instead of connections and waves.

Turn it on for one workflow and the others in your world keep their waves.

Place agents anywhere

• Drag a card to move it. It stays where you drop it.
• Drag several cards at once. Select them (Shift+click each one, or drag a box around them in Select mode), then drag any one of them. The others move with it.
• Add an agent. Drop a role from the palette where you want the card, right-click an empty spot and pick a role under Add Agent, or place one with Shift+1–9 and a click. The card lands on that spot. An agent added from a menu or the palette without a spot goes beside the others. A second agent of the same role gets a number (Developer Agent 2), because agents tell each other apart by name.
• Copy an agent. Duplicate (Ctrl+D) and paste (Ctrl+V) put each copy just below and to the right of the card it was copied from. Copy several cards and they keep how they sat to each other. A copy is called … (copy), and a second copy of the same card gets a number.
• Turn on "Needs review". The reviewer card goes beside the agent it reviews, on a free spot.

Groups

A group is a name and a light border around some of your agents, with one message box for all of them.

Make a group. Select two or more agents and press Group in the bar that appears at the bottom of the canvas, or press Ctrl+G. The new group is called Group 1, Group 2 and so on.

An agent is in one group at a time. Put it into a new group and it leaves the old one. A group with no agents left disappears.

Along the top of each group's border:

• The name. Click it to select the group's agents, then drag any one of them to move the whole group. Double-click it to rename the group.
• Who is in it. One small face per agent, in its role's colour, with its initials. Click a face to select that agent. A green dot on a face means that agent is replying. With a lead, the lead's face comes first, with a small crown. Past five agents, the rest show as +N.
• How many agents the group holds.
• Introductions (two speech bubbles) and Lead (a crown). Two switches, tinted while they are on. See Introductions and A lead for the group.
• The message box. See below.
• × removes the group. The agents stay where they are.

The border follows the cards: move a card and the border moves with it.

Message a whole group

Type in the group's Message box and press Enter or the send button. Each agent in the group gets the message in its own chat, as if you had typed it into that agent's box, and each one replies in its own chat. See Chatting with an agent.

The agents talk to each other on the board. Each agent is told the message went to the whole group, and who the others are. It is asked to introduce itself, say what it is taking on, share what it finds and read what the others post. They work at the same time, so each sees the others' notes as they are posted. The message in each chat says To the group and the group's name under it. See The board.

What happens for each agent is the same as when you message it yourself:

• An agent that is still replying gets the message as soon as its reply ends.
• An agent that can now do more than you last agreed to asks you to confirm first. Only one of these questions shows at a time. If another agent also needs to ask, its message is put back into its own chat box for you to send.
• When the message cannot go at all (for example, no model is connected, or your plan's active workflows are full), nothing is sent. The message stays in the group's box, with the reason under it, so you can send it again.

If the message did not simply reach every agent, a line under the box says what happened, for example Sent to 4 of 5. 1 was not sent; its chat says why.

Attach files and pictures. Press the paperclip at the left of the box, paste a file or a screenshot into it, or drop files onto it. Each one shows above your text, a picture as a small preview and any other file by its name. Press × on one to take it off. The files go with the message to each agent, exactly as if you had attached them in that agent's own chat, and they show in each chat. A message can be only files. If the message cannot go at all, the files stay in the box with your text.

Introductions

Press Introductions to turn it on for a group. From then on, each agent a message to the group reaches is first told: Start by introducing yourself to the others through the message board and collaborate on the user request.

Cerevisor adds that sentence to what each agent is told, the same way it tells them who else is in the group. Your message shows in each chat exactly as you typed it. When your plan does not include the board, the sentence is left out, because the agents cannot reach each other there.

A lead for the group

Press Lead to give the group a lead: an agent that coordinates the others for you.

• A new card joins the group, in front of the others, called after the group (Research lead, for example). It uses the workflow's model. Change its model or instructions like any other card.
• Agents with the same name get numbers. The lead tells agents apart by their names. If two agents of the group share a name, turning Lead on numbers them (Developer Agent, Developer Agent 2, Developer Agent 3), and the line under the box says which ones it renamed. One Undo puts the old names back and removes the lead.
• The message box now sends to the lead alone. Its placeholder reads Message Research through Research lead.
• The lead hands the work out. It reads your request, splits it into parts and gives each part to the agent that fits it best, in a note to that agent on the board. A note to an agent by name starts that agent's turn, so each one starts on its part. Each agent is told that the lead runs the work and that it should answer the lead on the board when its part is done.
• The lead can talk to the whole group at once. A note from the lead to anyone, or to no one in particular, starts every agent of its group, for something all of them need, such as "introduce yourselves" or a change to the plan. A note that only reports, such as the lead's "the file is built" at the end, starts nobody. A note to anyone from any other agent starts nobody either.
• The lead brings it together. Each answer starts the lead's next turn. When every part is in, it checks the parts against your request, asks for a fix where something is missing, and replies in its own chat with one answer, saying which agent did what. Open the lead's chat to follow the plan and read the answer.
• Your files reach whoever does the work. Files you attach go to the lead. Each agent the lead gives a part to gets the same files with the lead's first note to it, and sees them in its own chat. A later note does not send them again. Your next message to the lead brings its own files, or none.
• A message typed into the lead's own chat counts as a message to the group too, so the lead keeps coordinating.
• Agents can start each other more often. Between two of your messages, agents can usually start each other's turns 8 times. In a group with a lead the limit is 3 per agent of the group, up to 24.

Press Lead again to turn it off. Messages go to every agent again, and the lead card stays. Delete it if you no longer need it. Turn Lead on again and the same card leads again. If you deleted the card, a new one is added.

Without the board in your plan, the lead cannot reach the others, so it answers your request itself and says so.

Stopping agents

• One agent. While an agent replies, the Run agent button above its card is a red Stop. Press it to stop that reply.
• A group. While any agent of a group replies, a red Stop sits beside the group's send button. Its tooltip says how many are replying. Press it to stop all of them.
• Every agent of the workflow. On a free canvas the Run button in the Workflow bar is off, but while any agent replies it becomes a red Stop all agents. Press it to stop every reply on the canvas.

When you stop a group or every agent, messages that were waiting for a reply to end stay in each agent's chat. They are not sent by themselves. Send each with Send now when you are ready. Stopping one agent from its card works like the chat's own Stop: a message waiting behind the reply is sent next.

The AI Builder on a free canvas

The AI Builder knows when a workflow is on a free canvas, and works the way the canvas does:

• It adds agents without waves. Each new card goes beside the others, and you arrange them.
• It groups agents for you. Ask "put the researchers in a group called Research" or "add the editor to the Writers group". It can also rename a group or remove one.
• It does not wire agents or run the workflow. To ask one agent a question now, it sends the question to that agent's chat, and the answer appears there.
• It can switch modes when you ask. "Turn this into a free canvas" or "put these back into waves" works from the chat, the same as the button.
• It does not build step by step. Build and run step by step, in the builder's Setup, is greyed out on a free canvas, because it builds waves and runs them. If you had it on, it comes back on when you go back to waves.

Go back to waves

Press the Free canvas button again (its tooltip now reads Free canvas is on. Click to put the agents back into waves.). On a narrow canvas the row in More workflow actions reads Back to waves. Each agent goes into the wave nearest to where it sits, and each wave is stacked from top to bottom in the order its cards were. The lines between agents come back, and Run works again.

Your groups are kept for the next time you turn the free canvas on. They are only drawn on a free canvas.

Good to know

• Only single agents run on a free canvas. A chat message and Run this agent work as usual. A schedule, the Operator, a script or another workflow that tries to run the whole workflow is refused, with a sentence that says why.
• Undo (Ctrl+Z) undoes turning the free canvas on or off, making or renaming a group, ungrouping, and each of a group's switches. Undoing Lead also removes the card it added. Moving cards is not an undo step, as in waves.
• The setting and the groups are saved with the workflow, with their switches. A copy or a branch of the workflow keeps its groups, and each copied group is led by its own copy of the lead card.
• Sharing a workflow does not carry the free canvas or its groups. The person you share it with opens it in waves.