> ## Documentation Index > Fetch the complete documentation index at: https://docs.waterr.ai/llms.txt > Use this file to discover all available pages before exploring further. # Agent Computer > Give an agent its own cloud machine. It takes work from calls and chat, runs it after you approve, and hands back files. On a call, an agent can talk about the work. With an **Agent Computer**, it can also **do** the work. Each agent gets its own cloud machine. When someone on a call asks for something that takes longer than a conversation, such as research, a document, a spreadsheet or a follow-up, the agent writes it down as a **task**. Once you approve the task, the agent runs it on its computer, checks the result against what was asked and hands back the files. Nothing runs without a person deciding it should. Tasks from calls wait for your approval. Requests you make yourself in chat run straight away. > [!NOTE] > > Agent Computer is a **beta** feature, available on the **Enterprise plan**. A workspace admin has to turn it on before anyone in the workspace can use it. ## How it works **1. Someone asks for work** Someone asks for work on a call ("Can you put together a comparison of these three vendors after this?"), or you ask the agent in chat. The agent writes a task with a clear **requirement** and the **expectation criteria** that define done. **2. You approve it** Tasks from calls wait on the computer's page as **Assigned**. You read the task, edit it if needed, then **Approve** or **Reject** it. Requests you make in chat skip this step. **3. The agent works on its machine** The computer starts, the agent does the work in the task's project folder and the computer pauses again when the run ends. A run can last up to 90 minutes. **4. You read what came back** The task shows the agent's summary, its own check against each criterion and the files it produced. You can open, download, share or pass those files to another agent. ## Turn on Agent Computer Agent Computer is turned on for the whole workspace. Only workspace **owners and admins** can change it. **1. Open Settings › Labs** In the Studio, go to **Settings**, then the **Labs** tab. **2. Switch on Agent Computer** Under **Beta features**, turn on **Agent Computer**. This adds **Platform › Agent Computer** and an **Agent Computer** section to every agent's Settings. On plans below Enterprise the switch is locked. ## Set up a computer for an agent Each agent has one computer. **1. Open the agent's Settings** Open the agent in the Studio, go to its **Settings** tab and choose **Agent Computer**. **2. Click Set up a computer** Fill in the form: | Field | What to put | | - | - | | **Title** | A short name, such as "Sales ops desk" | | **Description** | What this computer is for. **The agent reads this**, so be specific: the kind of work, who it's for, any standards to follow. | | **Type** | **Work** for research, documents, browsing and follow-ups. **Coding** for writing and running code, with a repository per project. | **3. Click Turn on** Waterr creates the machine and runs a connection check right away. > [!TIP] > > Start with **Work** unless the agent's job is mainly writing code. Changing the type later starts the agent on a fresh machine, and files on the old one stay behind. You can't change the type while a task is running. Coding computers are still rolling out. If yours isn't available yet, its tasks say why they can't start. ### Check that it's ready The Agent Computer section shows an **On/Off** switch, **Test** and **Edit**. **Test** starts the machine, has the agent answer one request and pauses the machine again. You can watch the steps (**Machine → Agent → Reply**) with their timings. | Status | Meaning | | - | - | | **Connected · ready for tasks** | The machine and the agent both answered. Approved work will run. | | **Connected · running a task now** | A task is in progress | | **Connecting** | The machine is starting and the agent is being asked to answer | | **Not connected** | The check failed. The reason is shown underneath. | | **Not checked yet** | Click **Test** | | **Off** | The agent takes on no work | A test isn't charged. ## Where the work lives ### The computer's page Every computer has its own page with all of its tasks, projects and files. To open it, click **Open** on the **Tasks, projects and files** card in the agent's Agent Computer settings. It opens in a new tab. The page has: * **A status bar**: New → Assigned → In progress → Done, plus Rejected and Failed. Click a status to filter by it. * **Projects** on the left: **All tasks**, each project and **New project**. * **A task list** with each task's status, title, project, who asked and when it was written. * **The project's files**, when you pick a project. ### Platform › Agent Computer **Platform › Agent Computer** shows every computer in the workspace, with two tabs: * **Computers**: every agent's computer, with its status, task count, projects and last task. Click a row to open that computer's page. * **Tasks**: every task across all agents. Filter by **Waiting for approval** to see everything that needs a decision. The tab badge shows how many tasks are waiting. ## Tasks ### What a task contains | Field | What it is | | - | - | | **Title** | A one-line name | | **Requirement** | What to do, written for someone with no memory of the call | | **Expectation criteria** | The checks that define done, one per line. The agent tests its work against these. | | **Project** | Which project folder the work goes into | A task stays **New** until it has both a requirement and expectation criteria. If the agent couldn't get enough detail on the call, click **Edit** and fill in the gaps. ### Where tasks come from | Source | What happens | | - | - | | **The agent, during a call** | When someone on a call asks for work to be done afterwards, the agent writes the task. It waits on the computer's page as **Assigned** until you approve it. | | **You, in the agent's chat** | Ask the agent in **Chats** (or the agent's chat panel) to do something on its computer. If you can edit the agent, the task **starts right away**, because you're the one asking. | | **You, on a build call** | Start a video call with the agent from its chat. Tasks you agree on queue on the call's Notepad (see [Build calls](#build-calls)). | | **Apps using an API key** | Tasks requested through the API always wait for approval. | You can't write a task by hand on the computer's page. Ask the agent in chat and it writes the task for you. ### Statuses | Status | Meaning | | - | - | | **New** | The requirement or expectation criteria are still missing | | **Assigned** | Complete and waiting for your approval | | **Waiting** | Approved and queued to start | | **In progress** | Running on the computer | | **Done** | Finished. Read what came back. | | **Rejected** | Declined. It won't run and can't be approved later. | | **Failed** | It didn't finish. Click **Approve again** to retry. | ### Approve, edit or reject Open a task to see these actions: * **Edit**: change the title, requirement, criteria or project. Available while the task is New, Assigned or Failed. * **Approve**: run the task. On a failed task this button reads **Approve again**. * **Reject**: decline it, with an optional note on why. A rejected task can't be approved later. Tasks can only be approved from a signed-in browser session, not by the agent itself or by an API key. Before a run starts, Waterr checks that the computer is on and ready, that the workspace has credits and that a run slot is free. Each workspace runs up to **two tasks at once**. Any others wait their turn. ### What comes back When a run ends, the task shows: * **What came back**: the agent's summary of what it did * **Criteria report**: the agent's own check against each expectation criterion. This is the agent grading itself, so read the deliverables before you rely on it. * **Deliverables**: the files it published * **Run log**: the step-by-step record of the run For tasks started from chat, the result is also posted into that chat as **Done: …** or **Could not finish: …**, with the files attached. ## Projects Projects keep each stream of work in its own folder on the computer, so the agent always knows where to find earlier work. * Every computer starts with a **General** project, which is always there. * Click **New project**, give it a name and, optionally, a description of what goes in it. The agent picks the right project for each task. * Archive projects you've finished with (**Archive project**) and bring them back with **Restore project**. **Show archived** lists them. You can't archive the General project. ## Files Pick a project on the computer's page to see its files. Each file has: | Action | What it does | | - | - | | **Open** | View it in the browser: images, PDF, text, Markdown, CSV, JSON and logs | | **Download** | Save it to your device | | **Share** | Publish it as a public web page (see below) | | **Copy to…** | Send it to another agent's computer | ### Give the agent files to work with Use **Upload** in a project to add files, such as a price list, a brief or a data export, up to 100 MB each. They go into the project's `inputs/` folder, and the agent finds them there on its next task in that project. ### Files the agent didn't publish The agent publishes its final deliverables. To see everything else it changed, open **On the computer** to browse the project's folder on the machine and click **Get** to pull a file off. Files the agent changed but didn't publish also appear under **Also changed on the computer — not published**. ### Copy a file to another agent **Copy to…** sends a file to another agent in your workspace that has a computer. Pick the agent and the project, and the file lands in that project's `inputs/` folder, ready for its next task. This is how you hand work from one agent to another, for example research from one agent to the agent that writes proposals. ### Share a file on the web **Share** opens **Publish to web**: **1. Create link** Click **Create link**. The file is published as a page on `files.waterr.ai` that anyone with the link can view, without signing in. **2. Copy and send it** Click **Copy** and send the link. The dialog shows how many times the page has been viewed. **3. Update it when the file changes** The link is a snapshot. If the agent changes the file later, the dialog offers **Update link**, which replaces the content at the same address. **4. Stop sharing when you're done** **Stop sharing** turns the link off for everyone. Only a signed-in person can share a file. The agent can't publish anything to the web by itself. Shared files can be up to 50 MB. ## Use it from chat The quickest way to give an agent work is to ask it. Open **Chats** in the Studio, pick the agent and type what you need: > *"Pull the last quarter's pricing from the uploaded sheet and write a one-page summary for the sales team."* The agent writes the task and starts it. A task card appears in the chat and updates as it goes: **Starting the computer… → Working… → Done**. When it's finished, the card shows the files. You can **Preview** text files right in the chat, **Download** them, or click **Details** to open the full task on the computer's page. If the card says **Needs more detail**, the agent couldn't write a complete task. Reply with what's missing. ### Build calls Some work is easier to explain out loud. The **video button** next to the chat box (**Video call** with the agent) opens a call in a new tab, where you can talk the work through. * Tasks you agree on appear under **Computer tasks** on the call's **Notepad** as **Queued**. * Click **Send** to run a task now, while you're still on the call. When it finishes, the agent tells you. * Click **×** to remove a task that hasn't started. * Anything still queued **runs by itself when the call ends**, and the results show up in the agent's chat. A build call is charged like any other call, and the tasks are charged as computer runs. ## The agent can improve itself While it works, the computer can update its own agent's **draft**: the script, description, welcome message, persona and goals. Bigger changes, such as adding agents to the roster or changing tools, skills or guardrails, come to you as **suggestions** in the agent's chat with **Apply** and **Dismiss** buttons. The computer only ever edits the draft, never the live agent, and every change is labelled **Agent computer** in the chat so you can review or revert it. Nothing reaches your callers until you publish. ## Billing Each finished run is charged from your workspace credits in two parts, both listed under **Tools & more** in Billing: * **Agent Computer (sandbox)**: the time the machine was running * **Agent Computer (AI)**: the AI model's work during the run Connection tests aren't charged. If the workspace is out of credits, new work is refused with *"This workspace has used all of its credits. Top up before giving the agent more work."* ## Limits | Limit | Value | | - | - | | Computers per agent | 1 | | Longest run | 90 minutes | | Runs at the same time | 2 per workspace | | Upload size | 100 MB per file | | Shared file size | 50 MB | Agent Computer files are kept according to your workspace's data retention settings, which include an **Agent Computer files** window. ## Troubleshooting > [accordion] I don't see Agent Computer in Settings > > A workspace admin needs to turn on **Agent Computer** in **Settings › Labs**, and the workspace needs to be on the Enterprise plan. > [accordion] The status says Not connected > > Click **Test** again. The reason for the failure is shown under the status. If it keeps failing, turn the computer off and on, or contact support with the error text. > [accordion] A task is stuck on New > > The task is missing a requirement or expectation criteria. Click **Edit**, fill them in and the task moves to **Assigned**. > [accordion] A task failed > > Open it and read the **Run log** and **What came back** to see where it stopped. A common cause is a task too large for one 90-minute run. Edit the task to narrow it, or split it, then click **Approve again**. > [accordion] The agent's result mentions a file I can't find > > The agent may have saved it without publishing it. Open **On the computer** for that project and click **Get** on the file. > [accordion] Work from a call never showed up > > The agent only writes a task when someone asks for work to be done **after** the call. Check the computer's page, or **Platform › Agent Computer › Tasks**, filtered by **Waiting for approval**. ## Related * [Agent Mail](https://docs.waterr.ai/agent-tools/agent-mail.md): give the agent its own email address and calendar invitations * [Custom Functions](https://docs.waterr.ai/capabilities/custom-functions.md): let the agent call your own APIs during a call * [Knowledge Base](https://docs.waterr.ai/capabilities/knowledge-base.md): files the agent can look things up in during a call This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.