A sub-agent is Bluedeer itself, started a second time for one part of your request. It has the same tools and abilities as Bluedeer in your chat and follows the same rules. It gets a description of its task, works through it on its own with its own steps, and sends back what it found or did. Bluedeer then uses that to write its reply.
Which plans have them
Sub-agents are part of the Pro and Max plans, and of a shared workspace's seats on those plans. The Free and Plus plans do not have them: there, Bluedeer does the same work itself in the reply. Sub-agents are being introduced gradually, so they may not be available on your account yet even on a plan that includes them.
Why Bluedeer uses them
A sub-agent works like Bluedeer itself, so Bluedeer can delegate any self-contained part of a task to it, much as you would hand a task to a colleague. Research across many pages is one example. Others are writing a document, building a page or a design, working through an app, or analysing a set of files. The sub-agent does the work in its own space and sends back only its result, so its in-between steps, such as every page it read, do not stay in the chat. The chat stays smaller, so later messages in it stay quick and use less of your allowance.
Bluedeer decides when to use a sub-agent. It usually does so for a larger piece of work, or for separate parts of a task that can be done side by side, such as a design and the document that goes with it, or the rules of several countries compared one country each. For a quick question it answers directly. When a task has separate parts, Bluedeer can start several sub-agents at once, and they work at the same time.
One reply can start up to four sub-agents. Within that limit, Bluedeer decides how many to start. When a reply has already started four, Bluedeer does any further part itself, and that step says so.
Names and colours
Each sub-agent has a name that Bluedeer gives it, such as davinci or ada, and its own colour. No two sub-agents in the same chat share a name, and a new one gets a colour that no other sub-agent of the chat has, so you can tell them apart at a glance. The name and colour stay with the sub-agent when you come back to the chat later.
What a sub-agent starts with
Bluedeer starts each sub-agent in one of two ways:
- From the conversation so far. The sub-agent starts with a copy of the chat, so it knows the context of your request without Bluedeer repeating it.
- As a new conversation. The sub-agent starts with only its task, together with your instructions, the chat's project and your memories, as Bluedeer itself would have them.
Either way, Bluedeer writes the task so that it is complete on its own, including what the sub-agent must keep to from your chat, such as your standing instructions on language, tone or format and what you have already decided. A sub-agent that is asked to read the web normally starts as a new conversation, so the pages it reads do not sit next to the files, emails or earlier messages from your chat.
A sub-agent thinks as much as the reply it belongs to, following the thinking mode of your chat, unless Bluedeer asks it to think more or less for its particular task. See Which thinking mode should I use?.
What you see
Each sub-agent appears as one line in the thinking section above the reply, where its step is. The line shows the sub-agent's mark in its colour, its name and how it is going:
- Queued when it waits for another sub-agent of the same reply to complete. Its mark is an empty ring.
- The step it is on and how many steps it has taken, such as Reading a page · 4 steps, while it works. Its mark moves while it works, or stays still if your device is set to reduce motion.
- Completed, with the number of steps it took, when it has completed its task.
- Limit reached, with the number of steps it took, when it reached the most steps it may take for one task. It then sends back what it did and what it did not finish, and Bluedeer carries on from there.
- Not completed, with a warning mark, when it could not complete its task. Bluedeer then carries on without it, usually by doing that part itself in the same reply.
- Stopped when you stopped the reply.
Press a line to open it. Once the reply is complete, you see the task Bluedeer gave the sub-agent. You also see each step the sub-agent took, such as Searched the web or Read a page, with what it searched for or read. A step that read web pages has a sources link, such as 3 sources, which lists the pages it read. When the sub-agent has finished, Result shows what it sent back to Bluedeer. A long task or result shows its first lines with Show all. When a sub-agent could not complete its task, the reason appears when you open its line, for example when the AI provider did not respond or when the available usage had been used up.
When several sub-agents run at once, each has its own line with its own name and colour, in the order Bluedeer started them. They are not combined into one line like other steps taken at the same time. See What is the thinking section above a reply?.
The lines stay with the reply. When you come back to the chat later, or reload the page while Bluedeer is still working, you see the same lines, names, colours and steps.
What a sub-agent can and cannot do
A sub-agent can do what Bluedeer can do in your chat, under the same rules. It can search the web and read pages, read and search your artifacts and earlier chats, use apps you have connected, make and change artifacts, pages, documents, designs, charts and presentations, save memories, use your computer, and set up scheduled tasks. The only thing it cannot do is start sub-agents of its own.
What a sub-agent makes appears in the reply as it would if Bluedeer had made it, for example an artifact card or a design. A small line above such a card shows the sub-agent's mark and says which sub-agent made it, such as Made by davinci. The cards stay with the reply and open as any other card does.
Anything that needs your approval or an answer from you still asks first. When a sub-agent wants to send an email, share or publish something, delete an artifact, or anything else that asks first, the approval card appears in the reply, as it would for Bluedeer. The same goes for a question to you, such as a request to connect an app or to fill in a secure field. The sub-agent does not wait for your answer: it finishes what it can and tells Bluedeer what is waiting. When you answer the card, your answer goes to Bluedeer, which carries on from it. A sub-agent never gets more access than Bluedeer has in the same chat, and nothing it asks for is approved for you. See What are action confirmations and approval cards?.
On your computer, each sub-agent works in its own folder. It is told which other sub-agents of the reply are using the computer at the same time, and it takes care not to touch their work. When two sub-agents of one reply change the same file, they take turns, so neither overwrites the other.
A sub-agent works only with what you could use yourself in the workspace you are in. In a chat that others can read, it uses only what everyone who can read the chat may see, as Bluedeer does. See Who can write in a shared chat?.
Usage
A sub-agent's work counts toward your usage in the same way as the reply it belongs to. A sub-agent that starts from the conversation so far reads it at the lower price for text the AI provider has recently seen. Because the chat keeps only each sub-agent's result, the messages you send afterwards in that chat usually cost less than if Bluedeer had taken every step in the chat itself.
What others see in a shared chat
Everyone who can read the chat sees each sub-agent as you do: its name and colour, its state, its steps, the sources it read, the task it was given and what it found. The only thing kept to the person who sent the message is what each step searched for or read, as for Bluedeer's own steps. The cards a sub-agent made show to everyone with the same Made by line.
When Bluedeer support looks at a chat at your request, it sees each sub-agent's line and plain steps, but not its name, its sources, its task or what it found.
Related
- What is the thinking section above a reply? explains the other steps and the thinking time.
- Which thinking mode should I use? sets how much Bluedeer thinks before it answers.