Knowledge base / Research and tools

Team mode

For a big research question or a large change to a repository, Nymbot can split the work between several models at once, with your own model in charge.

What is Team mode

In Team mode, your pinned Pro model becomes the overseer. It splits the task, hands the parts to two to four workers, and puts their work together. The workers all run on one model that you choose, which can be a cheaper one than the overseer.

Team mode is for two kinds of work only: deep research and tasks in a connected repository. A message that is neither is not sent; turn Research on or connect a repository, or send it without Team mode.

Without a Pro model pinned to act as the overseer, Team mode is refused and nothing is sent.

Team mode is set per chat. With a Pro model pinned, a Team chip appears in the chat's toolbar whenever the next message would be deep research (the Research chip is on) or a repository task (a repository is connected to the chat). It is off until you set it up. Tap it to open the Team mode sheet:

• Number of workers: 2, 3 or 4. The default is 3.
• Worker model: the model every worker runs on. The app suggests a cheaper, faster one, the lowest-priced metered model from the same maker as your pinned model when there is one. Tap Change to pick another from the model list.
• A price line that updates as you change either one: Up to N Pro credits · usually about M. The first number is the most the message can cost, and the second is what a team like this usually costs.

Tap Save and the chip lights up and reads Team of 3 (or however many workers you chose). From then on, research messages and repository tasks in that chat are sent to a team. Open the sheet again and tap Turn off to go back to a single model. If the chat has connectors or server runs on, the sheet says the overseer can use them and that each call waits for you to allow it.

A team on deep research

The overseer breaks the question into sub-questions and gives each worker its own share. The workers research them in parallel, each running a few searches of its own and reading what they find.

The overseer then reconciles what they found and checks where they contradict each other. If one sub-question came back weak, it can send that one back to be researched again, once. Then it writes the report itself, with sources, as in an ordinary research report.

A team in a repository

The overseer first looks around the repository, reading without changing anything. It then splits the change into parts, each with its own files, so no two workers touch the same file. A worker can only list, read and search the repository and edit or write its own files; it cannot commit, open branches or pull requests, or use anything else.

When the workers are done, the overseer reviews the combined change. If one part is not right, it can send that part back to be redone, once. A worker that fails has its edits rolled back.

From there it is the same as any repository task: the overseer commits the usual way, or the changes are held for you to review first if the repository asks before committing.

Connectors and server runs

The overseer can use the chat's connectors and, in a repository task with server runs on, run commands on a Nymbot server. It can do this while it plans, while it reviews the workers' change, and before it splits a research question. Workers never get connectors or server runs.

Every connector call and every server run the overseer wants stops the team and shows the usual approval card in the reply, marked as coming from the team lead. This happens every time, even for a connector tool you set to Always allow in other chats, so that card has no Always allow this tool button. Allow it and the team carries on from exactly where it stopped; deny or decline it and the overseer is told and carries on without it. Either way, the workers' finished work is kept, nothing is run twice, and you are not charged again for what already ran.

A server run is charged on its own, for the time it actually runs, on top of the team's cost. When a chat has a spending cap, a run is only offered if it fits in what the cap leaves after the team's maximum.

Letting the overseer use tools adds a few of its steps to the maximum shown on the Team mode sheet.

What it costs

A team costs the metered total of every model involved: the overseer plus every worker, each metered at its own model's Pro price. Before you send, the app shows the most the task can cost: the overseer's budget plus each worker's. That maximum is held from your Pro balance, you are charged only for what was actually used, and the rest is released.

• If your Pro balance cannot cover the maximum, nothing is sent.
• If the maximum is more than a chat's spending cap allows for one reply, nothing is sent and nothing is charged.
• Each worker has its own budget. One that reaches it stops and hands back what it has so far.
• A worker that fails is not charged, and a team task that fails is not charged at all.
• A long task can pause and carry on in another step, within the same maximum.

If a chat has a spending cap, the app checks the team's maximum against it before sending, and asks first if it could go over, just as it does for any other message. When a team is refused, the chat says Team mode did not start and why, gives your message back, and nothing is charged.

Watching the team work

While a team works, the reply shows a lane for the overseer, labeled Lead, and one for each worker, labeled Worker 1, Worker 2 and so on, so you can follow what each of them is doing: searching, reading, editing, finished, stopped at its own budget, or failed.

If the model provider starts limiting how fast the team can go, the remaining workers run one at a time instead of all at once, which takes longer.

Under the finished reply, a short Team summary lists the Lead and each worker, with the model it ran on, the Pro credits it used, and whether it was done, stopped or failed. It also says when the workers had to run one at a time. The price on the reply is the total for the whole team; tap it to see the breakdown, with a Team lead row and one row for each worker.

The Tasks pane

The checklist button at the top of the chat opens its Tasks pane: beside the chat on a wide screen, as a sheet on a phone. While a team works, the button shows it is running, and the pane lays the work out as an outline, newest turn first: the Lead's plan, each worker's part with the model it runs on, its cost and whether it is done, stopped or failed, then the Lead's review and final steps. A connector call or server run the Lead asks for is listed with its approval, and you can allow or decline it right there, exactly as on the card in the chat. Every request still running is listed at the top with its own plan, Stop and Add instructions, including ones started on your other devices. Tap any step to jump to its place in the chat. The outline is kept with the chat's messages, so it is still there after the work finishes and on your other devices.