← Back to Blog
Claude Code Agents Explained: Subagents, Agent View, and Teams.
Claude Code has four ways to run agents in parallel. Subagents, agent view, agent teams, and worktrees each do something different. Here is how to choose.
Claude Code has four ways to run agents in parallel. Subagents, agent view, agent teams, and worktrees each do something different. Here is how to choose.
Claude Code has four ways to run agents in parallel: subagents, agent view, agent teams, and worktrees. Each one handles a different coordination problem. Subagents offload a side task and return a summary. Agent view runs background sessions you monitor from one screen. Agent teams let multiple agents share a task list and message each other. Worktrees isolate parallel sessions so they never overwrite each other's files.
The right one depends on three questions: who coordinates the work, do the workers need to talk to each other, and do the tasks touch the same files.
I'm Tom. I teach Claude Code operators how to run production workflows, and I went through the full Anthropic agent documentation to give you the clean version without the noise.
In Claude Code, an agent is a Claude session with its own context window, its own set of tools, and its own instructions. When Claude delegates work to an agent, that agent runs the task independently and returns only a result, not its entire working context.
That is the key difference from asking Claude to do something in your main session. An agent's context stays separate. Your main conversation does not fill up with search results or file reads the agent needed to do its job.
Anthropic's documentation lays out the trade-offs in a single table, and it is worth understanding before you reach for any specific approach.
Subagents are delegated workers inside one session. They do a side task in their own context and return a summary. Use them when a task would flood your main conversation with content you will not reference again.
Agent view is one screen for dispatching and monitoring background sessions. You open it with claude agents. Each background session runs without a terminal attached, so you can close agent view and the sessions keep working. This is a research preview and requires Claude Code v2.1.139 or later.
Agent teams are multiple coordinated sessions with a shared task list and inter-agent messaging. Agents can message each other directly without routing everything through you. This is experimental and disabled by default.
Worktrees are isolated git checkouts. Parallel sessions each get their own copy of the repository so they can edit without touching each other's files. Agent view creates worktrees automatically for background sessions that edit files.
There is also /batch, which splits a large mechanical change into 5 to 30 worktree-isolated subagents, each of which opens its own pull request. It is the right tool for repo-wide migrations or mechanical refactors you can describe in one instruction.
The decision comes down to three questions.
Who coordinates the work? If you want Claude to delegate and collect results inside one conversation, use subagents. If you are handing off independent tasks and checking back on them, use agent view. If you want Claude to plan, assign, and supervise a group of workers that coordinate among themselves, use agent teams.
Do the workers need to talk to each other? Subagents report results back to the conversation that spawned them. Agent view sessions report only to you. Teammates in an agent team share a task list and message each other directly.
Do the tasks touch the same files? If yes, use worktrees. Agent view does this automatically for background sessions. Subagents and sessions you run yourself can each get their own worktree too.
Subagents are the most mature and most used approach. Each one runs with a custom system prompt, specific tool access, and independent permissions. You can also control costs by routing subagent tasks to a faster, cheaper model like Haiku instead of Opus.
Custom subagents live in your user config or in a project's .claude/ directory. Claude uses each subagent's description to decide when to delegate. Write a clear description and Claude starts routing tasks automatically.
The /agents command (inside a session) opens a panel showing your running subagents and lets you create or edit custom ones. This is separate from claude agents, which opens agent view. Both use the word "agents" but they do different things.
To track in-session background tasks, use /tasks.
Agent view is the right tool when you have several independent tasks that Claude can run without you watching every step. Open it with claude agents in your shell, dispatch a task from the bottom input, and each task becomes a row showing its state.
Sessions stay alive even when agent view is closed. A background supervisor process keeps them running, so you can dispatch work, close the terminal, and check back when you are ready. Sessions survive sleep but not shutdown.
Row summaries are generated by a Haiku-class model. They refresh every 15 seconds while a session works, and once when each turn ends.
Agent teams let agents message each other and share a task list instead of reporting everything through you. A lead agent creates the team, creates tasks, and spawns teammates. Teammates claim tasks, message each other when they need to coordinate, and mark work done.
This is experimental and disabled by default. Enable it by adding CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to your settings.json env block. You need v2.1.32 or later.
Agent teams do not isolate teammates in worktrees, so partition the work so each teammate owns a different set of files. Otherwise two agents editing the same file will conflict.
For more on agent teams setup and real examples, read Claude Agent Teams: How to Run Multiple AI Agents at Once.
Worktrees give each parallel session its own git checkout under .claude/worktrees/. Sessions can read the same codebase but each writes to its own copy.
You do not always need to manage this yourself. Agent view creates a worktree automatically before a background session edits files. For manual parallel work, you can create worktrees yourself and start Claude sessions in each one.
The Agent SDK is for building agents that run outside Claude Code, via the Anthropic API. If you are building a production agent you will deploy as a service, that is the right tool. For agents that run inside your Claude Code workflow, subagents and agent teams are the right abstraction. The SDK and Claude Code agents serve different audiences.
For a deep look at the SDK, read Claude Agent SDK: Build Your First Production Agent.
Subagents run inside your current session and return a result when done. Background agents (started through agent view or claude --bg) are separate sessions that keep running without your terminal attached. You can check on them from agent view and attach whenever you want.
No extra charge, but they multiply token usage. Each subagent has its own context window, and running several at once burns through your quota faster than a single session would. You can reduce the cost by assigning subagent tasks to Haiku instead of Opus in the subagent's frontmatter.
There is no hard limit, but rate limits apply. Running many background sessions in parallel uses your subscription quota proportionally. Anthropic's documentation notes that ten sessions in parallel uses quota roughly ten times as fast as one.
Agent view is in research preview as of May 2026. It requires Claude Code v2.1.139 or later. The interface and keyboard shortcuts may change. Agent teams are even earlier, still experimental and off by default.
Start with subagents. They are the most stable, most documented, and easiest to reason about. Once you are comfortable delegating tasks to subagents, move to agent view for background monitoring, then consider agent teams only when you need agents to coordinate with each other.
The operators who get the most from Claude Code are not the ones who run the longest single sessions. They are the ones who delegate well, keep their main context clean, and let agents handle the work that would otherwise flood the conversation.
Subagents are where most people should start. Agent view is where you go when you want to stop watching. Agent teams are where you go when the work itself needs coordination.
Grab the free Claude Code Blueprint for the foundations, or join the 30-day challenge to build real systems with agents in your business.
Five interactive lessons. Install Claude Code, build your first automation, and deploy it live on the internet — all in under an hour. Free, no coding required.
Grab the Blueprint →