Work with Agents

Open Grove from your Team’s App menu to talk to an Agent. A Session is one conversation with one Agent. Starting another Session gives you a separate conversation, not another Agent or another Computer.

Start a conversation

1. In Grove, choose Create.
2. To make an Agent, enter a name under New Agent, choose New Computer or No Computer, and choose Create Agent. To use an existing Agent, select it under Start a conversation with.
3. Choose Start, then send your first message.

If Grove asks for Agent access, review its permissions. It needs permission to list Agents and to read and send their messages. Creating an Agent requires separate permission. An Agent can exist before Grove has conversation access to it.

Give the Agent a goal, the relevant files, and a way to check its work. Put task-specific instructions in your message. Don’t paste passwords or API keys into the conversation. See model keys for credential setup and billing for funded work.

Illustrative: this prompt describes a file task, not a guaranteed model result. Use a new folder on the Agent’s Home Computer.

TEXT

Create notes/hello.md with three weekend project ideas. Read the file back
and tell me its path. Ask before replacing any existing file.

Return to saved work

Choose a conversation from Grove’s inbox to reopen its Session. All Agents shows the Team’s visible conversations. Choose an Agent in the filter to narrow the list. The inbox separates conversations waiting for you, work in progress, and finished or failed work.

You can switch conversations while another Agent works. If a Session cannot be opened, Grove keeps the previous conversation open. Try opening the failed Session again. Retry conversations reloads an unavailable inbox.

Sending is not finishing

An accepted message can wait behind work already running. Queue sends a follow-up for later. Cancel stops the current run. A waiting message has its own Cancel action, which withdraws that message if it has not started. Cancelling does not undo files or other changes already made.

Read the final reply and check the result before treating the task as done. Open Inkcap, select the Agent’s Home Computer, and inspect a file the Agent wrote. Open Shed on that Computer to inspect processes or run a check.

Models and failed runs

Model choices depend on the host and the credentials or funded models available to your Team. A saved model preference does not prove the provider is ready. If the conversation offers model selection, choose from its available catalog.

For a failed or interrupted run, read the notice before choosing Retry. Retry sends the failed prompt again. Check for partial changes first, especially before repeating a command that creates or sends something. Model work already done can use credits even when the run fails or is cancelled.

To work in parallel, use separate Sessions and be explicit about which files each task may change. Sessions with the same Agent share its Home Computer. They do not isolate edits to the same file.