Commands, settings, and limits for running Claude Code work in parallel, for someone at the keyboard deciding how to split a task across subagents, sessions, teams, or a project.
Delegate inside one session
- Delegate a side task — ask in plain words: "Use a subagent to …". It runs in its own context and returns a summary.
- Use a built-in —
Explore(read-only search),Plan(plan-mode research),general-purpose(multi-step work). - Fork with full context —
/subtask <task>starts a subagent that inherits the whole conversation and prompt cache. - Define a specialist — Markdown file in
.claude/agents/(project) or~/.claude/agents/(all projects); onlynameanddescriptionare required. - Call a specialist — type
@and pick it, or run the whole session as it withclaude --agent <name>. - Check running subagents —
/taskslists them with status and model; attach to or stop any one. - Turn fork mode off —
CLAUDE_CODE_FORK_SUBAGENT=0lets Claude wait on subagents in the foreground again.
Run sessions in the background
- Open agent view —
claude agentsshows every background session and which ones need you. - Dispatch from a shell —
claude --bg "<task>"; add--name <name>to label it. - Background the current session —
/bg(alias/background);/forkcopies the session into a new background session. - Peek, attach, detach — Space peeks, Enter attaches, ← on an empty prompt detaches. Detaching never stops a session.
- Manage from a shell —
claude attach <id>,claude logs <id>,claude stop <id>,claude agents --json.
Hand work to a coordinator
- Start a project — claude.ai/code or the desktop Code tab → Projects → New project. Public beta, Pro and Max.
- Set the brief once — Project settings › Memory › Project instructions, up to 16,000 characters, reaches every new thread.
- Steer the coordinator — say "propose threads before starting" or "run at most two threads at a time" in the project conversation.
- Run a thread locally —
claude remote-controlin the folder (v2.1.280+), then choose Work locally from the + menu.
Message between sessions
- Name a session —
/rename <name>orclaude --name <name>; peers address it by that name. - See who's reachable —
/list-agents(alias/peers); the first line is this session's own name. - Send a message — "Let @<name> know …". Claude writes it and delivers with SendMessage.
- Wait for another session — "Tell me when <session> finishes" gets one notice when it next goes idle (v2.1.236+).
- Control inbound —
crossSessionInbound:accept,hold, orrefuse; set it in/configunder Messages from your other sessions. - Gate cross-machine sends —
"isolatePeerMachines": trueasks before any message leaves this machine.
Run an agent team
- Enable teams — set
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1in env or settingsenv; off by default. - Spawn teammates — "Spawn three teammates: one on UX, one on architecture, one devil's advocate."
- Split panes —
"teammateMode": "auto"uses tmux or iTerm2 when available; default is in-process.
Scale with workflows
- Ask for a workflow — say "use a workflow" or include
ultracodein the prompt. - Research across sources —
/deep-research <question>runs the bundled cross-checking workflow. - Watch or save a run —
/workflows;ppauses,xstops,ssaves the script as a slash command.
Gotchas
/forkchanged meaning — with agent view on (v2.1.212+),/forkmakes a new background session; the in-session fork is/subtask./agentsis notclaude agents—/agentspoints to subagent files;claude agentsopens agent view.- Teams hijack named subagents — with agent teams enabled, a subagent Claude names launches as a teammate.
- Teams share one checkout — teammates don't get worktrees; give each a different set of files.
- Cloud threads skip repo config — project threads load CLAUDE.md but not repo skills,
.mcp.json, hooks, or settings.json. - Coordinator limits are advice — thread caps told to a project's coordinator are preferences, not enforced limits.
- Messages can't grant permission — a cross-session message never approves a prompt or changes another session's config.
- Parallel work burns plan limits — every subagent, teammate, thread, and workflow agent spends its own tokens.
