Skip to main content
An agent session holds a task’s conversation and tool activity. Start one for each job and name the repository or working directory in the first message. Choose New agent, then pick an online runtime in the composer. A local path must exist on that runtime’s computer.

​
Select tools and credentials

Select the connections needed for the task. A selected team connection is still subject to the member allow-list and provider permissions. Personal credentials belong to their member; inviting someone into a session does not give them your personal accounts. Some external entry points have different initial credential selection, so review their setup guide before starting work there. Mention a connected resource in the prompt when you want the agent to inspect that exact account, cluster, deployment or issue. Resource context helps resolve names; it does not silently grant an unselected credential.

​
Share a session

Use the session’s Share control to review its access. You can also @mention a teammate and choose Invite. Check whether you are sharing read access or permission to contribute. Participants see the same transcript and can steer work only when their session and computer permissions allow it. Sharing the conversation and sharing its computer are separate. A teammate who cannot use the machine may be able to read without sending work there. Review computer sharing before inviting someone to a local runtime: its existing sign-ins are powerful. To end access, remove the participant or change the session’s general access in its sharing controls. Removing a person from one conversation is not the same as revoking their access to the computer or connector.

​
Steer, answer and stop

Tool calls and results appear in the transcript. Send a correction during a turn to steer it at the next supported boundary. Answer question cards in Needs your input or the Inbox. Stop the running turn when you need it to halt; stopping is not rollback of commands already executed. Permission modes determine runtime action prompts. Plans cover significant cloud infrastructure work, not every code edit or ticket. Say explicitly when you want a diff, commit or publication review before it happens.

​
Move work to another runtime

Use the conversation’s runtime picker and review the move confirmation. Prepare the destination repository, branch, dependencies and required sign-ins. A different computer does not automatically have the files on the old one. Keep changes in git or transfer the required files, then ask the agent to check its working directory before continuing. History provides task context; a move may continue with a different provider session rather than the original native CLI session. Keep machine and provider differences in mind.

​
Continue an existing CLI session

For supported local runtimes, start a new conversation, choose +, then Continue a Claude Code session or Continue a Codex session. Pick the session available on that computer and review the import confirmation. Long imports may omit older turns. Importing history does not clone its repository.

​
Search and return later

Use workspace search to find prior conversations. Archive finished agents to reduce clutter; find archived work under the team’s Settings → Archived agents. An archived transcript is different from a deleted computer. History is persisted and long context may be compacted. A remote or cloud runtime can keep working independently of your laptop, but local execution needs the computer and runtime running. A saved transcript does not mean a background process or watcher survived a restart.

​
Other entry points

Slack and other connected messaging channels can continue agent work in their own threads. Scheduled and webhook work uses triggers. Their initial permission mode and credential selection can differ from an interactive Desktop session; inspect those settings before using production.