Skip to main content

Work with a cloud agent

You launch and manage cloud agents in the Fluence Console under Cloud agents > Agents. For turns, idle stop, and pricing, see Cloud agents concepts.

Launch an agent​

Before the first launch, connect an AI provider and add a GitHub token. See Launch your first cloud agent.

  1. On the Agents tab, click New agent.
  2. If you have more than one preset, choose the Preset. With only the Default preset, the form does not show this field.
  3. Optional: enter an Agent name, up to 128 bytes on one line.
  4. In Task, describe the work. The task is the first message the agent gets. It can be up to 20,000 characters.
  5. In Model, choose a model.
  6. In Repository URL, enter an https:// address without credentials, for example https://github.com/owner/repository.
  7. Check the Review panel and click Launch agent.

When you choose another preset, the form takes its default model and fills the task and the repository if you left them empty. The Machine step shows the machine of the preset. To change it, click Change in preset.

Repositories​

A launch needs a repository. The agent clones it with the GitHub token:

  • The token needs read access to clone, and write access to push and open pull requests.
  • The Repository step shows Token checked, Token not checked, or No GitHub token. A checked token is valid, but the console learns only at launch whether it reaches this repository.
  • An agent that cannot clone the repository fails. Its page shows The repository could not be prepared on the agent's machine. Check the address and the access of the token, then launch again.

An agent works on one repository. To work on another one, launch another agent.

Agent list​

The list shows each agent with its repository as the title, its Last turn, Model, Price per hour, Updated time, and State.

  • Last turn shows Running, OK, or the reason the last turn went badly.
  • By default, the list shows active agents. Ended and failed agents are hidden. To see them, change the state filter, or click Show all.
  • Search agents matches the repository, the agent name, and the agent ID.
  • To sort by the last change, click Updated.

The list refreshes by itself every 10 seconds.

Agent page​

Click an agent to open its page. It has the conversation on the left and the Details panel on the right.

The Details panel shows the Status, Name, Model, Repository, Machine, the number of Extensions the agent was launched with, and the Price of the machine per hour. If the agent failed, the panel says why.

The page refreshes by itself every 5 seconds while the agent is starting, working, or stopping.

Send a message​

Type in the field under the conversation. A message can be up to 20,000 characters. What happens depends on the state of the agent:

Agent stateWhat you can do
IdlePress Enter or click Send. The message starts a new turn at once.
WorkingClick Queue or press Enter to run the message as its own turn after the running one. Click Steer now or press Ctrl+Enter (⌘+Enter on macOS) to hand it to the running turn at its next step.
Pending, PreparingWait. The agent takes messages once it is Working or Idle.
Stopping, Ended, FailedThe agent no longer takes messages.

Press Shift+Enter for a new line.

Use Steer now to correct the agent while it works, for example "Do not touch the migrations." Use Queue for the next piece of work. A queued message runs as a separate turn with its own time limit.

Delivery labels​

Each message that you send shows a label:

LabelMeaning
QueuedAccepted, waiting for its turn
SendingBeing handed to the agent
DeliveredThe agent has it
unknown — checkingFluence is checking whether the agent got it. Wait. Do not send it again, or the agent can get it twice.
FailedIt did not run. The reason is shown next to it.
CancelledIt never ran. It was still waiting when the agent was stopped or ended.

If a running turn is cut short, a message that you steered into it and that did not run shows Failed. A queued message is kept and runs as the next turn.

Interrupt a turn​

To stop what the agent is doing and keep the agent, interrupt the turn:

  1. On the agent page, open the menu in the Details panel.
  2. Click Interrupt agent. The menu shows it only while the agent is Working.

The running turn stops, and the agent becomes Idle. It keeps its machine and its conversation, and it takes your next message.

Stop an agent​

  1. Open the menu in the Details panel of the agent page, or the menu on the agent's row in the list.
  2. Click Stop, and confirm with Stop agent.

The agent stops, and its machine is released. Messages that have not run yet will not run. You cannot start a stopped agent again. Anything the agent did not push to the repository is gone with the machine.

An idle agent also stops by itself after the idle time of its preset. See Idle stop.

Launch a copy​

To start a new agent with the settings of an existing one, open the menu of the agent and click Launch a copy. The launch form opens with the name, task, model, and repository of that agent, and with its preset if the preset still exists. Change what you need, and click Launch agent.

The copy is a new agent on a new machine. It does not continue the conversation of the first one.