A started session can only use accounts of its own provider. Delegation is how work crosses providers: a session hands a task to a child agent on another account or provider on the same machine. A Claude session can ask Codex, Gemini on Antigravity, an OpenCode model, or a model on a second Claude account.
A child agent is a real BeyDock session. Its transcript, approvals and usage are visible. It doesn't appear in the sidebar. You open it from the parent's transcript. Child agents can't start agents of their own.
Turn it on
Delegation is on by default. The switch is Settings › Agents › Let sessions start agents, and the same switch is under Home › Connections › BeyDock. When it is off, sessions don't get the agent tools.
Ask for an agent
Tell the agent in plain words, for example "have Codex write the tests while you refactor". You can also type @ in the message box and pick a target from the Agents section, which lists provider, account, model and effort. BeyDock adds a note to your message with the choice. BeyDock doesn't choose models for the agent. The agent, your instructions or a skill you turn on decide that.
Before an agent starts, an approval card shows the whole task, the target, whether it can edit, and where it will work. Settings › Agents › Ask before starting an agent controls when this appears: Always, Only for Agents That Edit (the default) or Never.
Access: read-only or edit
- Read-only is the default. The agent works in the session's folder and can't change anything. Several can run in parallel.
- Edit runs the agent in its own git worktree. Its changes come back as a diff. They reach your folder only through Apply to the folder, which asks you to confirm, or when the parent agent applies them and you approve the "Apply agent changes" card. If there are conflicts, nothing is applied.
Choose accounts
An agent can name an account or ask for auto. Auto picks the account with the lowest five-hour usage that offers the model, and skips accounts above their ceiling. Naming an account always works.
Under Settings › Agents you control:
- Accounts agents can use and which models each account may run, and Account order: when an agent names no account, the first signed-in account in this order is used (auto uses it only to break ties).
- Limits: agents per session (3 by default) and per device (4 by default).
- Usage ceilings: the five-hour (85 percent) and weekly (95 percent) defaults, with per-account overrides.
- Agent tools in sessions: turn the tools off for a provider or an account.
- Delegation skill: off by default. It teaches agents when and how to hand work off. Duplicate it as your own skill or write a new one.
Follow the work
Each agent shows as a row in the parent's transcript with its status. From the row you can open its session, stop it, or keep it in the sidebar. When it finishes, its report goes back to the parent agent. The parent treats that report as data from another model, not as instructions.
If you archive or delete a session that has running agents, BeyDock asks whether to stop them first.
Provider notes
- All four providers can start agents and run as agents.
- Antigravity can't ask for approval mid-turn, so an Antigravity agent's permissions come from its mode. Stop on the agent's row interrupts it like any other agent.
- OpenCode's free models are checked regularly. A model that stops responding is hidden from agents until a later check succeeds. See Free model health in Settings › Agents.
- Agents run on the same machine as the parent. On an SSH host, that is the host.