# How to Run Multiple Claude Code Agents at Once > Running several Claude Code agents at once works if each has its own copy of the project and you can see them all. How it works and where it goes wrong. - Canonical: https://orcstrator.com/run-multiple-claude-code-agents-at-once - Published: 2026-11-04 - Author: Rodrigo Nask (https://rodrigonask.com/), OrcStrator --- ## Key takeaways - Several agents can work at once, but each needs its own copy of the files or two will overwrite each other. - A board of cards shows which agent is working, finished or waiting on you. - More agents spend your usage limit faster, so watch the quota while they run. - Quiet chats should stop running, and come back with the conversation cache still warm. Running multiple Claude Code agents at once means several Claude Code sessions working on different jobs at the same time. It works safely when each agent has its own copy of the project and you can see which ones need you. You gave Claude Code a job: rewrite the pricing page. It will be busy for a while. Next week's emails still need drafting, and a third job is waiting behind both. So you sit and watch one chat, because starting a second one feels like asking for trouble. The feeling is right, and the answer is still yes. Several agents can work at the same time. What decides whether it goes well is not how many you start. It is whether they can step on each other, and whether you can tell which one is waiting for you. > **Figure (animated on the web version):** Five rows, each a different Claude Code agent with its own job: rewrite the pricing page, draft next week's emails, compare three supplier quotes, fix the broken contact form and summarise the meeting notes. Their progress bars all fill at the same time at different speeds, and the contact form row finishes first and turns green. > > Five jobs, five agents, one moment. **They work side by side**, each on a different job, and report back when they are done. ## Can Claude Code run more than one agent at once? Yes. You can have several Claude Code sessions open at the same time, each working on its own job. In this guide, an agent means one Claude Code session: one conversation, doing one job, with its own memory of what it was told. Claude Code offers two ways to do it, and they are easy to mix up. The first is separate sessions. You start another one, give it another job, and both run. Anthropic's documentation describes this under [parallel sessions with worktrees](https://code.claude.com/docs/en/common-workflows): work on a feature in one terminal while Claude fixes a bug in another, without the edits colliding. The second is helpers inside one session. A session can send out [subagents](https://code.claude.com/docs/en/sub-agents), which are smaller assistants that each work in their own context and report back a summary. They suit research. They are not a team of independent workers on separate jobs. The documentation itself points to separate sessions for larger parallel work. Claude Code runs in a terminal, which is a text window where you type commands. If that is not your world, the second half of the problem is the harder one: starting a second agent is easy, but keeping track of five is not. The rest of this guide covers both halves, using OrcStrator, a Windows app that [runs Claude Code for you as chats on a board](https://orcstrator.com/how-it-works). OrcStrator uses Claude Code and is not affiliated with Anthropic. ## What goes wrong with two agents on one project Three things go wrong, and they have nothing to do with how smart the agents are. **They overwrite each other.** Two agents pointed at the same folder are two people editing one paper page. One changes the headline while the other rewrites the paragraph under it. Whoever saves last wins, and the other agent's work disappears without a warning. **You cannot see them.** Each agent lives in its own window. Five agents means five windows, and one of them stopped an hour ago to ask a question. Nothing on your screen told you. The agent did the polite thing and waited, and you lost the hour. **They drink from one cup.** Every working agent draws on the same Claude plan allowance. Five at once use it up faster than one. You find out halfway through the job, when Claude says you have reached your limit. Overwriting, blindness and quota. The first is the only one that can destroy work, so it comes first. ## Why each agent needs its own copy of the files The fix for overwriting is old and plain: give every agent its own copy of the project to work in. Developers call this a git worktree. Git is the tool most software projects use to keep a history of every change. A worktree is a second working copy of the same project, in its own folder, on its own branch. A branch is a separate line of changes that stays apart from the main version until someone merges it back. The [git worktree manual](https://git-scm.com/docs/git-worktree) describes it as a way to have more than one working tree attached to a single repository. Claude Code can start a session in one, and its documentation describes the result as a separate checkout on its own branch. The picture is two people each holding their own photocopy of the page. They can both scribble. Nothing clashes until someone decides to put the copies back together, and at that point a person looks at the result. > **Figure (animated on the web version):** Two panels. On the left, one shared page with the headline edited by agent A and agent B at the same time, ending in a red mark and the words Work overwritten. On the right, two separate copies of the page, one for each agent, each keeping its own edit, ending with a green line that says Each copy kept its own work. > > Shared files mean overwritten work. **A separate copy per agent** keeps each edit apart until you decide to bring them together. In OrcStrator you do not type any of that. A chat can work on its own copy of the project, though it does not by default. A chat that does has "Open its working copy" in its right-click menu. When you close a chat, Safe Close in its Close menu asks the chat to save only its own work, then closes it. The app also checks what a chat leaves behind when you close it, and the Close menu reads "Nothing left behind" when there is nothing loose. By default chats share the project folder, so there is a second layer. OrcStrator watches which files each chat has changed and not yet saved. If a second chat in the same folder reaches for a file that the first is still changing, the second one is paused, and can be resumed, with a note saying why, instead of quietly overwriting. ![One chat edits the page while a second chat waits, because it knows the first one has unsaved changes there](https://orcstrator.com/features/aware.webp) Caption: Two chats, one page. **The second waits** because it knows the first has unsaved changes there. A separate copy stops accidents. It does not decide whether the work is good, so you still read what each agent made before you accept it. ## How a grid shows every agent at a glance The answer to blindness is putting every agent on one screen. Plain Claude Code gives you a window per session. You tab between them and remember which one was doing what. OrcStrator shows your chats side by side in a grid, one tile each, so you read the whole situation in one look: which are working, which stopped, how much of its memory each chat has used. ![Three Claude chats side by side, each working on a different job for the bakery](https://orcstrator.com/features/interface.webp) Caption: Three chats side by side. **Each tile is one agent** on one job, with its progress visible without opening anything. A tile has a header with the chat name, the project it belongs to, and what the chat has cost so far. A dot on the left shows its state. If you have too many tiles, you can hide one without stopping it: the chat keeps running, and the tile comes back when you want it. The tile that matters most is the loud one. A chat that is waiting on you, because it asked a question, wants a plan approved or needs a permission, gets an amber header, an amber outline and a "Needs you" label next to its name. A chat that finished and has not been read gets a thin amber line along the top. A chat whose tool has said nothing for ten minutes gets a red "Stuck on" label with the tool name and how long it has been silent. > **Figure (animated on the web version):** A grid of six chat tiles with green working dots. After a moment the Weekly emails tile turns amber with the words Needs you and Asks a question, then the Contact form tile turns amber with Needs you and Wants a permission. The other four tiles stay green and keep working. > > Amber means it needs you. **A question, a plan or a permission** turns the tile amber and says why, so you never scan six windows for the one that stopped. That is the whole idea. You do not scan five windows for the one that stopped. The screen tells you which tile is asking. ## How a board of cards keeps the jobs straight A grid shows what is running now. A board shows the jobs themselves, including the ones nobody has started. In OrcStrator every job is a card, and every chat belongs to one card. Cards move across a board from Backlog to Ready, Running, Needs your review and Done. You write a job on a card in plain words, press Start, and an agent picks it up. When it finishes, the card moves to Needs your review and waits for your check. ![The board with cards in columns from Backlog to Done](https://orcstrator.com/features/board.webp) Caption: The board. **Every job is a card**, and the column it sits in tells you where it stands. You are not juggling five conversations. You are managing five cards, each with its own chat and history. Which job stalled on Monday? The answer is a column, not a memory. There is also a dashboard that gathers what needs you in one list: chats that asked a question, cards in Needs your review, and anything stuck. It is the screen to open first in the morning. If you would rather ask, The Orc, the assistant built into OrcStrator (it comes with the top plan, not the free one), answers in plain words when you ask what is stuck or what finished while you were out. If you want to see the board in use, the home page shows [a board with five agents working](https://orcstrator.com/#inside). If you are starting from nothing, starting your first card takes you through one job from the first sentence to Done. For how several sessions stay straight, keeping several sessions straight covers the same ground from the session side, and a task board for Claude Code explains why a board suits people who do not code. ## Which agents should wait for you? Not every agent needs you, and the ones that do should be easy to spot. There are three reasons an agent stops and waits. **A question.** The agent reached a choice it will not make for you: which of two layouts, which tone, which customer list. It asks, and stops. **A plan.** For bigger jobs you can have an agent write out what it intends to do before it does any of it. It stops until you approve. This is the cheapest place to catch a wrong turn, because nothing has changed yet. **A permission.** Claude Code asks before it does something with side effects, such as running a command or changing files. In a terminal that is a line of text you can miss. In OrcStrator it is a question in plain words, and the safe requests that keep coming back can be allowed once instead of answered all day. The risky ones still stop and ask. All three turn the card amber and say which one it is. That is the whole signal. The other direction matters as much. An agent that is working should be left alone. The temptation with five cards is to read each one every few minutes. Do not. Look at the board when something turns amber, answer, and go back to what you were doing. Your attention is the scarce part of this setup, not the agents. ## What parallel agents do to your usage limit Every working agent draws on the same allowance, so more agents means a faster drain. Your Claude plan has a session limit and a weekly limit. Anthropic's [help page on usage limits](https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work) explains that the session period starts when you send your first message and resets in full when it ends, and that weekly limits apply across all models. It also says usage is affected by the length and complexity of your conversations, and that your use of all Claude product surfaces counts toward the same limit. Check that page for the current figures, because they change. Nothing in that arrangement gives each agent its own allowance. An agent working on a long, large job spends more than one on a short job. Several working at the same moment spend it together. Helpers inside a session count too: the subagents documentation says each one sends its own requests, which count toward the same usage limits as the main conversation. > **Figure (animated on the web version):** An illustration of one shared allowance with two bars. The first bar, labelled 1 agent working, drains slowly. The second bar, labelled 5 agents working, drains much faster and turns red. Both draw on one session and one week. > > More agents spend the limit faster. **Five working agents draw on one session and one week**, so the same allowance runs out sooner. The bars are an illustration, not a measurement. OrcStrator cannot make the allowance bigger, and it does not try to. It does four things to make it last longer and to keep you informed. - **It shows the limit on screen.** Bars for the session and the week sit at the top of the sidebar. When your pace would use it all up within the next few hours and before the window resets, a line says so, for example "Out of quota in 2h 13m", while there is still time to do something. - **It keeps chats cheap to come back to.** A large chat left alone is compacted, meaning shrunk into a summary, before Claude forgets it, so the next message does not reread everything. Settings has a switch to keep chats warm for an hour. - **It sends the boring work to cheaper helpers.** Searching and reading go to a smaller helper model by default, so they stop drawing on the main one. - **It lets each chat load only what it needs.** Tools, skills and plugins switch on and off per chat. ![Quota bars for the session and the week, with how long until you run out](https://orcstrator.com/features/quota.webp) Caption: The session and the week. **The line under the bars** tells you how long you have at the current pace. For how the limit works in more depth, read how the usage limit works. The short version for this guide: the number of agents you can run is mostly decided by this allowance, not by your computer. ## How many agents is too many? Fewer than you think, and fewer than the screen invites. The first limit is the allowance, which the quota bars show within a day or two. OrcStrator's free plan also caps how many chats can work at the same time. The number lives in the plan and can change, so read it in Settings, under Account & updates, in "Your OrcStrator account", and not from this page. Settings, under Chats & models, also has presets for how many chats work at once. The second limit is you. Every agent that finishes lands in your Needs your review column, and every one needs a decision. Five agents that finish at once is five decisions at once. If that column grows faster than you empty it, you have too many running. The sign is not the machine slowing down. It is a pile of finished work you have not read. The third limit is the job. Two agents are the right number when their jobs touch different things: a page and an email, a report and a proposal. They are the wrong number when the second job depends on the first, or when both would change the same file. Then one agent working in order is faster, because the second would only wait or redo its work. So start with two. Give them jobs that do not overlap, each in its own copy. Watch the board and the quota for a day. Add a third when the second was easy to review and the bars still look healthy. The machine is rarely the limit. In plain Claude Code, a terminal you leave open stays alive and holds some of your computer's memory. OrcStrator closes a chat's process the moment its turn ends and starts it again from the saved conversation when you write next. On the home page the example is twenty chats open and two working, which means two programs running on your computer. The other eighteen are still there, with their conversation saved and the cache still warm, and a message brings one back in about a second. > **Figure (animated on the web version):** Twenty small circles for twenty open chats. Two glow purple and are labelled working. The other eighteen are grey and labelled waiting, not running. A message arrives for one grey circle and it lights up purple, with the note back in about a second. > > Quiet chats stop running. **The conversation waits, the computer does not.** A message brings a quiet chat back from its saved conversation in about a second. One catch: your computer has to be on for your agents to work. They run on your machine, on your own Claude account. ## Frequently Asked Questions ### Can I run several Claude Code sessions on the same project? Yes, but give each its own copy of the files. Two sessions editing the same folder can overwrite each other's work. A separate working copy per session, which developers call a worktree, prevents it. OrcStrator can run a chat in its own copy, and Safe Close in the chat's Close menu saves that chat's own work before closing, so you never touch git. ### How many Claude Code agents can run at once? The limit is mostly your Claude plan's usage, not the software. Each working agent draws on the same weekly and session limit, so five agents spend it faster than one. OrcStrator's free plan caps how many chats work at the same time, and the number is shown in Settings. ### Do idle agents use my computer's memory? In plain Claude Code, a terminal you leave open stays alive. OrcStrator closes a chat's process when the turn ends and starts it again from the saved conversation on your next message, in about a second. Twenty chats open with two working means two programs using your computer. ### How do I know which agent needs me? You watch the screen, not the terminals. In OrcStrator every chat is a card on a board and the ones waiting on a question, a plan or a permission turn amber and say why. The Orc, the built-in assistant on the top plan, can also list what is stuck when you ask in plain words. ### Is there a risk that agents undo each other's work? Yes, when they share files. Separate working copies remove most of it. By default chats share the project folder, and OrcStrator pauses a second chat that reaches for a file another chat is still changing. You still review the results before accepting them. ## Run your agents on one board OrcStrator puts every agent on one screen: a grid of chats, a board of cards, amber when one needs you, and your usage limit on screen while they work. It is free to start, it runs only on Windows 10 and 11, and it needs your own paid Claude plan (Pro, Max, Ultra, Team or Enterprise, per [Anthropic's setup page](https://code.claude.com/docs/en/setup)). Read next: Claude Code Multiple Sessions Without the Chaos, Claude Code Task Board for Non-Coders and Claude Code Git Worktrees Explained Plainly. Put your jobs on a board, give each agent its own copy of the project, and see which one needs you without opening five windows. - [Download OrcStrator](https://orcstrator.com/#install) If you are new to all of this, Claude Code for non-coders is the place to start. For the mechanics of copies in more detail, read git worktrees in plain words, or [download OrcStrator for Windows](https://orcstrator.com/#install) and put two agents on two small jobs today.