Claude Code parallel agents: which ones Claude uses on its own, and which you must ask for

TL;DR

• Claude uses these on its own: subagents, forked subagents, background bash, cross-session messages, and threads inside a Project.
• You must start these: agent view, /batch, routines, and workflows (unless you turn on /effort ultracode).
• Agent teams are off by default. Once enabled, Claude can form a team on its own, and it does not ask first.
• To force a feature: @-mention the subagent, type ultracode, or run its slash command. To stop Claude using one, change a setting or add a deny rule.

Legend

Mark	Meaning
🤖 Auto	Claude decides to use it without being asked
👤 Manual	Only runs when you run a command, flag, keyword or UI action
🔀 Both	Claude may use it on its own, and you can also ask for it
⚙️ Off by default	Must be enabled in settings before either side can use it

The five main approaches

Feature	Trigger	What Claude does on its own	How you trigger it yourself	How to stop or limit it
Subagents	🔀 Both	Picks a subagent from its description. Writing "use proactively" in the description makes it delegate more often	Name it in plain language ("use the code-reviewer subagent"). Claude usually follows, but it's still Claude's call. @-mention (@agent-<name>) always runs it. claude --agent <name> or "agent": "<name>" in settings runs the whole session as that agent	Remove or narrow the subagent's description. Deny the Agent tool
Agent view (background sessions)	👤 Manual	Nothing. Claude never starts a background session. Once you start one, it moves itself into a worktree before it edits files	claude agents (then type a prompt), claude --bg "task", /background (/bg), /fork	"disableAgentView": true or CLAUDE_CODE_DISABLE_AGENT_VIEW=1. Worktree step: "worktree": {"bgIsolation": "none"}
Agent teams	⚙️ then 🔀	Once enabled, any subagent Claude gives a name becomes a teammate. Teams can form when you never asked for one, and Claude Code doesn't ask first	Enable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, then say "spawn 3 teammates to…". Works only in interactive sessions, not -p or the SDK	Set the variable to 0. Claude then spawns ordinary subagents again
Projects (claude.ai/code, desktop)	👤 create, then 🤖 threads	Starts a cloud thread for each task you post, or passes the task to a thread already working in that area	You create the project (New project). A thread runs on your own computer (via Remote Control) only when you ask	Tell Claude to propose threads before starting them, to limit how many run at once, or to "answer here instead of starting a thread". Claude saves these to project memory
Dynamic workflows	👤 Manual by default, 🤖 with ultracode	Nothing, unless /effort ultracode is on. Then Claude plans a workflow for each substantive task, sometimes several in a row	Type the ultracode keyword or say "use a workflow". Run /deep-research <q> or a saved /<name>. The keyword only works in prompts you type, not in -p, schedules or webhooks	/effort ultracode off. Turn off the "Ultracode keyword trigger" in /config. Turn workflows off with "disableWorkflows": true or CLAUDE_CODE_DISABLE_WORKFLOWS=1

Supporting tools

Feature	Trigger	Automatic part	Manual part
Worktrees	🔀 Both	Background sessions from agent view or --bg move into a worktree before editing. Subagents with isolation: worktree in their frontmatter always get one. Claude can also give a fork its own worktree	claude -w <name> (or -w "#1234" for a PR). Ask Claude to "work in a worktree", which uses the EnterWorktree tool. Ask it to "use worktrees for your agents". Desktop worktree option. git worktree add
Cross-session messaging	🔀 Both (on by default)	Claude messages your other sessions without being asked, for example to warn them about a breaking change	Ask "tell the session in my other terminal…". @-mention a session (@api-worker). Run /list-agents (/peers) to see which sessions it can reach. To stop it: crossSessionInbound: refuse, or deny SendMessage and ListAgents
/batch	👤 Manual	Once you run it: plans 5–30 units, waits for your approval, then spawns one background subagent per unit, each in its own worktree	/batch <instruction>. The docs don't say Claude invokes it on its own, and it isn't in the list of skills Claude can call itself, so treat it as user-only

Related features that aren't parallel agents

Feature	Trigger	Detail
Background bash	🔀 Both	Claude can run a long command in the background itself. You can ask it to, or press Ctrl+B on a running command (twice in tmux)
Forked subagent	🔀 Both	Fork mode is on by default in interactive sessions, so Claude can spawn forks (subagents that keep the full conversation) itself. You start one with /subtask <task>. Control it with CLAUDE_CODE_FORK_SUBAGENT=0/1
/fork	👤 Manual	Copies the whole session into a new background session. If agent view is off, /fork makes a forked subagent instead
Background subagents	🤖 Auto	Where fork mode is on, every subagent Claude spawns runs in the background. CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1 makes them run in the foreground
Routines	👤 set up, then 🤖 runs	You create them with /schedule (/routines), at claude.ai/code/routines, or in the desktop app. After that, they run by themselves on a schedule, an API call or a GitHub event. Claude doesn't create one unless you ask, which you can also do inside a Project

Watching running work (all manual)

Command	Shows
claude agents	Agent view: every background session and its state
/tasks (/bashes)	Background work in this session: bash commands and subagents, including finished ones
/workflows	Workflow runs. p pauses, x stops, s saves the run as a command
@ typeahead	Named background subagents and their status
/agents	Only prints where subagent files live. It is not the same as claude agents
/list-agents	Sessions, subagents and teammates that Claude can message

Watch for these

• Agent teams change ordinary delegation. With the flag on, a subagent Claude names becomes a teammate, which costs more tokens. Turn the flag off if you only want subagents.
• Ultracode applies to every task in the session. Expect more tokens and slower answers. It also skips the "Large workflow" warning and the subagent cap.
• Asking for a subagent by name is a strong hint, not a guarantee. Use @agent-<name> when it must run.
• Fork mode is on by default. Claude may spawn background forks without being asked. Interactive sessions behave differently from -p and the SDK here.

Sources

• Run agents in parallel
• Subagents · Agent view · Agent teams
• Dynamic workflows · Projects · Routines
• Worktrees · Cross-session messaging
• Commands · Skills · Interactive mode

Checked against the docs on 2026-10-05.