Four Ways to Run Parallel Agents in Claude Code – Picking the Right Approach

Subagents: temporary workers within a session

Subagents are child agents that Claude launches inside the current conversation. Each subagent has its own context window and returns only a summary, leaving its search results, logs, and file contents isolated from the main session. Built‑in subagents include the read‑only Explore (runs on Haiku, cheap and fast), Plan for planning, and a general‑purpose agent. Users can also create custom subagents with the /agents command, specifying allowed tools and model. Subagents cannot spawn further subagents and only report back to the session that created them.

Agent view: dispatch and check back later

When you have several unrelated tasks and prefer not to monitor each step, use the Agent view. The command claude agents (still a research preview) opens a panel where each background session occupies a line, showing its status, errors, and completion. Each session runs as a full Claude Code dialogue, independent of the terminal, and can be inspected by pressing space or entered for interaction. Agent view automatically places each session in its own Git worktree, providing file‑level isolation, unlike Subagents which report directly to the user.

Agent teams: Claude as leader with teammates

Agent teams let Claude act as a leader coordinating a group of teammate agents. Unlike the previous methods, teammates can communicate directly, share a task list, claim tasks, and exchange messages. This feature is disabled by default and must be enabled by setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json. The recommended team size is three to five members; larger teams increase coordination overhead and token consumption because each teammate runs an independent Claude instance. Agent teams do not provide automatic worktree isolation, so users must manually assign separate files to each teammate. Combining Agent teams with tmux is recommended for effective multitasking.

Dynamic workflows: script‑driven orchestration

Dynamic workflows replace Claude’s round‑by‑round decision making with a JavaScript script that schedules subagents at scale. The script, generated by Claude, can launch dozens to hundreds of agents (runtime limit: up to 16 concurrent, 1,000 total per run). Benefits include massive parallelism—useful for whole‑repo bug scans or large migrations—and higher quality, as the script can enforce cross‑validation among agents. The workflow is invoked with /workflows, also a research preview, and includes a built‑in /deep-research skill that produces a multi‑angle report with citations. Because such workflows can consume many tokens, the author advises cautious use.

Worktrees and /batch

Git worktrees solve file‑conflict isolation but do not coordinate work distribution. A worktree is created with claude --worktree feature-auth, giving an independent directory and branch while sharing the repository history. Worktrees can be combined with Subagents or Agent view for isolated parallel sessions. The /batch skill packages a series of subagents each with its own worktree, effectively automating the creation of multiple PRs, but it is not a distinct coordination model.

How to choose the right parallel‑agent method

Claude Code’s official guidance reduces the decision to three questions:

Who coordinates the work? Use Subagents if Claude handles delegation within a single session; use Agent view to dispatch and monitor later; use Agent teams when Claude should lead a collaborative team; use Dynamic workflows for script‑driven orchestration.

Do the agents need to communicate with each other? Only Agent teams provide direct teammate communication; Subagents and Agent view report solely to the creator.

Will the agents operate on overlapping files? If so, employ worktrees to isolate file changes. Subagents and independent sessions can each use a worktree, while Agent teams require manual file partitioning.

By answering these questions, users can select the most suitable parallel‑agent strategy for their specific workload.