A field guide for people who explain Claude Code
Where should this work flow?
Claude Code can put more than one Claude to work at once — but "run agents in parallel" is really four different features, and the right one depends on the shape of the work. This guide routes you by the question you're actually asking, not by feature name.
First, the picture
Think of a Claude Code session as a river: one channel with finite capacity — everything Claude reads flows through it, and the channel is only as big as the session's working memory. Every parallel-work feature is a different way of adding channels:
- Subagents — a side channel: a helper splits off, does a noisy side task in its own memory, and rejoins carrying only the answer.
- Agent view — the braids: several full sessions run in the background, and one screen shows which need attention.
- Agent teams — twin channels, one crew: a lead session runs teammates that share a task list and message each other. Experimental, off by default.
- Dynamic workflows — the canal: for big jobs, Claude writes a script that runs dozens to hundreds of workers, and the route can be rerun.
Two tools support all of the above without running agents: worktrees (levees — separate copies of the code so parallel work never floods the same files) and cross-session messaging (signals — one session passing a finding to another).
The routing questions
Start from your situation; the feature is the answer.
Test runs, log digs, doc fetches — side work that produces pages you'll never look at again. A subagent does it in its own memory and returns only the summary, so your conversation stays clear.
Hand each off as its own background session and watch them all from one screen. Agent view (opened with claude agents) shows what's running, what's done, and which session needs you. In research preview.
Reviewers with different lenses, investigators testing rival theories — agent teams give you a lead session plus teammates that share a task list and message each other. Experimental and off by default.
A 500-file migration, a whole-codebase audit, research worth cross-checking. A dynamic workflow moves the plan into a script that runs dozens to hundreds of workers in the background — and can verify their findings against each other.
Not if you separate them. A worktree gives each session or subagent its own copy of the code. Agent view does this automatically; agent teams don't — there, each teammate owns different files.
Cross-session messaging lets one session pass a finding, a status, or a warning to another — instead of you re-explaining it in the other terminal.
More Claudes means more usage — every approach multiplies it, against the same plan limits. The honest comparison is relative, not absolute.
The approaches, side by side
Chart No. 1 — Four ways to add channels
condensed from the official comparison table| Approach | What it gives you | Reach for it when | Who steers |
|---|---|---|---|
| Subagents | Helpers inside one session that do a side task privately and return a summary | A side task would flood your conversation with output you won't reread | Claude, inside one conversation |
| Agent view | One screen to dispatch and monitor background sessions. Research preview | Several independent tasks you want to hand off and check at a glance | You — hand off, check back later |
| Agent teams | Coordinated sessions with a shared task list and messaging, managed by a lead. Experimental, off by default | You want Claude to split a project, assign pieces, and keep workers in sync | A lead Claude session |
| Dynamic workflows | A rerunnable script, written by Claude, that runs many workers and can cross-check their results | A job outgrows a handful of helpers, or findings need verifying | A script holds the plan |
To see the difference in practice, One task, three ways runs the same PR review three ways — with where each one breaks down.
The supporting tools
Neither runs agents. They keep the channels from interfering.
A separate working copy of the code per session or helper, so parallel edits never collide. Needs a git repository.
Official docs →Sessions pass short text messages — never files or history — and a message can't approve anything or change the receiving session's settings.
Official docs →What it costs, honestly
Running several Claudes at once multiplies usage, against the same plan limits — the docs' own example: ten parallel sessions use quota roughly ten times as fast as one. The relative ordering they support:
- Subagents — the lower-cost option: results come back summarized, not as full separate conversations.
- Agent teams — costlier: each teammate is a separate Claude, so usage scales with team size. The docs' one figure: about 7× the tokens of a standard session when teammates run in plan mode.
- Workflows — "meaningfully more" than conversation; the docs' advice is to pilot a small slice first and watch usage live in
/workflows.
The docs publish no per-approach dollar figures, so neither do we. Their budget-level number — roughly $13 per developer per active day, enterprise average — covers all Claude Code use, not any one approach.