Can I have several Claudes working at once?

Yes. An agent team splits one request across several Claudes working in parallel — one per file, page, or question — and then gathers their findings into a single result. You start one by describing the goal in plain language, not by writing code.

Do I need to write JavaScript to use a dynamic workflow?

No. You describe the job in plain language and Claude writes the orchestration script for you. Knowing roughly how the patterns work — fan-out, verification, tournaments — just helps you ask for the right shape. You read and approve; you don't author the code.

When is a dynamic workflow overkill?

Most everyday tasks. A single rename, a quick fix, a normal review — one Claude in one context handles those fine, and a workflow would just burn extra tokens. Reach for a workflow when a job is genuinely large, highly parallel, or needs independent verification — not when it's merely big.

How is this different from the workflows in your other guide?

The other guide is about habits — planning, memory files, slash commands. A dynamic workflow is a specific feature where Claude writes a script that spawns and coordinates many subagents, each with its own clean context. One is how you work; the other is a tool you point at a hard problem.

What happens if I interrupt a workflow halfway through?

It resumes. If you quit the terminal or stop a run, picking the session back up lets the workflow continue from where it left off rather than starting over — the finished steps are cached.

Can a workflow use subagents inside it?

Yes. A workflow fans out work across agents, and each of those agents may spin up its own subagents for research tasks within their piece. They compose naturally.

When is a workflow overkill?

When the task doesn't split cleanly into independent pieces. If step 2 depends on the result of step 1, a workflow won't help — just use the main agent or a single subagent.

How many agents can run in parallel?

There's no hard limit you need to worry about — Claude manages the concurrency. The practical limit is whether the task splits into truly independent pieces.

Agent teams and parallel work — Claude for Humans

Most work with Claude is one conversation doing one thing at a time, and that is the right shape for most work. It stops being the right shape when a job is wide — forty pages to check, twelve files to migrate, six questions to research independently. Done in sequence that is an afternoon; done in parallel it is a few minutes, and each worker gets its own clean context instead of one conversation slowly filling up with forty pages of notes.

The pattern underneath is always the same: something splits the goal into independent pieces, each piece runs on its own, and the results are combined at the end. The important word is independent. Parallelism buys you nothing when step two needs the answer from step one — that is a sequence wearing a team costume, and it will be slower and messier than just asking once.

Two distinctions cause almost all the confusion here, and each has its own page below. A subagent is one helper sent off on a side-quest; an agent team is many helpers dividing one job. And a workflow is the heavier version of a team, where Claude writes a small orchestration script for the specific job in front of it rather than fanning out ad hoc. You do not write that script — you describe the job and read what comes back.

Starting any of this is a plain-language request, and it works the same in Claude Desktop as in the terminal. The management side is more technical: /workflows opens the live view of running fan-outs, and the claude agents screen shows every session you have going at once. Those are Power-Track tools, and you will not want them until you are genuinely juggling several jobs — at which point you will want them badly.

Worth knowing that the same instinct operates a level below teams, automatically. When Claude has several steps that don’t depend on each other — read these four unrelated files, run these two searches — it fires them together instead of waiting for each in turn. One wait instead of four. It’s why Claude sometimes does five things in what feels like a single beat, and you never have to ask for it. The reason to know it’s happening is that it tells you how to help: requests whose parts don’t secretly hinge on one another can run together, while a chain where every step needs the previous one’s answer is forced back into a line no matter how it’s phrased.

One habit worth forming early: ask for a single combined report at the end. Ten agents will happily hand you ten answers, and reading ten answers is the work you were trying to avoid.

The words

Multi-agent system Multi-agent orchestration · Agent orchestration: A multi-agent system is several AI agents working on one job at the same time, each handling a piece, instead of a single agent doing everything in sequence. Usually one 'lead' agent breaks the goal into parts, hands each to a teammate or subagent, and combines the results. You get breadth — many angles explored at once — and isolation, since each agent keeps its own clean context window.
Worktree git worktree: A second, real checkout of the same repository in its own folder, on its own branch. Not a copy you made by hand — git knows about it, both folders share one history, and switching branches in one doesn't disturb the other.