What developers run in agent sandboxes
Workloads that need an isolated computer for an AI agent, and how to build each one.
- Explore many solutions in parallel with sandbox forksFork one prepared sandbox into up to ten copies, run a different agent attempt in each, score them alike, keep the best.
- How to add a code interpreter to a chatbotGive each conversation its own sandbox with a stateful interpreter, run the model's code there, and show the text, tables and charts.
- How to build a browser-based online IDE on cloud sandboxesBack each workspace with a sandbox: files API for the editor, a terminal over WebSocket, live file events, and an app preview.
- How to build a coding interview platform with sandboxesPrepare a snapshot for each question, start a sandbox from it for each interview, run the candidate's code there, and record every edit.
- How to build a data analysis agent that runs Python safelyUpload the user's data to a sandbox, run the model's pandas code in a stateful interpreter there, and return tables and charts.
- How to build a Lovable-style AI app builder on sandboxesGive every project a sandbox from a template snapshot, let the model edit it through tools, preview it live, and publish on a domain.
- How to build a text-to-SQL agent that queries your database from a sandboxRun the model's SQL from a sandbox that can reach only your database, over a read-only login with a statement timeout, and return the rows.
- How to build an AI code review bot that runs the code it reviewsCheck the pull request out in a sandbox, let the model run its tests and linters there, then post the review from your service.
- How to convert Word documents to PDF on a server with LibreOffice headlessInstall LibreOffice in an isolated sandbox image, upload the file, run soffice --headless --convert-to pdf, and download the PDF.
- How to extract text from PDFs, scanned ones included, in a sandboxOpen each PDF in a throwaway sandbox, pull the embedded text with Poppler or pypdf, and run Tesseract OCR on pages that are only images.
- How to generate charts from LLM-written code in a sandboxRun the model's plotting code in a sandbox's code interpreter and take each chart back as a PNG result, or as an SVG or PDF file it saved.
- How to generate unit tests with an LLM and keep only the ones that passRun each model-written test in a sandbox three times and keep it only if it always passes and raises coverage.
- How to give an AI agent memory that persists between sessionsKeep each user's agent in a named sandbox that pauses between sessions, and put memory that must outlive it on a volume.
- How to give an AI research agent a sandbox to fetch pages and run codeGive the agent one sandbox per research question, with fetch and search tools inside it, a shell for analysis and files for its notes.
- How to give an LLM a code execution tool with function callingDeclare a run-command tool in your model's function-calling format, run each call in a sandbox, and return exit code and output.
- How to give every hackathon team a cloud dev environmentCreate one named sandbox per team from a starter image, give teams SSH and a public demo link, cap the spend, and stop them all by label.
- How to give every user their own cloud dev environmentGive each user a named sandbox that pauses when idle and wakes on the next request, with their files, memory and processes kept in between.
- How to grade student code safely in a sandboxRun each submission in its own throwaway microVM with the internet off and a time limit, run your tests against it, and record the score.
- How to have an AI agent set up any repository in a sandboxClone the repo into a fresh microVM, let the agent install and test until the suite passes, then snapshot the machine for later tasks.
- How to host MCP servers for your agents in sandboxesStart MCP servers inside a sandbox, give your agent each server's Streamable HTTP URL and bearer token, and keep API keys out of it.
- How to keep an AI agent running for hours in a cloud sandboxRun the agent in a sandbox whose lease renews itself, start it with spawn, and reattach to its output from any process later.
- How to let a Discord bot run code in a sandboxDefer the interaction within 3 seconds, run the code in a throwaway microVM with no network, then edit the original reply with the output.
- How to let a Slack bot run code safelyAcknowledge Slack within 3 seconds, run the snippet in a microVM with no internet and a time limit, then post the output to response_url.
- How to let an AI agent analyze Excel spreadsheets safelyUpload the workbook to a sandbox with the internet off, edit it with pandas and openpyxl in a stateful interpreter, and download it.
- How to let an AI agent test your own API in a sandboxRun your API and the agent's tests in one microVM, allow only your own hosts, and send each failing test back to the agent to fix or report.
- How to let an AI agent upgrade dependencies and prove nothing brokeInstall the project in a sandbox, fork one copy per outdated package, test each upgrade alone, and fix only the failures.
- How to preview apps an AI agent buildsRun the agent's app in its sandbox, give its port an HTTPS preview address, and let the sandbox pause between visits and wake on the next.
- How to reward an RL policy by running its code in sandboxesRun each sampled program in an offline sandbox, feed it every test input with a time limit, and score the share of outputs that match.
- How to run a browser automation agent in a cloud sandboxGive the agent a Linux desktop in a sandbox, send it screenshots, and apply the clicks and typing it chooses.
- How to run a coding playground for your usersRun each visitor's code in its own microVM sandbox with the internet off, a time limit per run, and a warm interpreter between runs.
- How to run an agent's CI on every pull requestUpload each pull request's code to a fresh sandbox, run install and tests there with no secrets inside, and report the exit code.
- How to run an AI code migration safely in sandboxesRecord which tests pass on the old toolchain, let the agent port the code on the new one, and accept a port only if none of them fail.
- How to run background AI agents in cloud sandboxesTake each task off a queue, give it a sandbox named after the task, run the agent there, and collect the result when it exits.
- How to run end-to-end tests in parallel cloud sandboxesPrepare one sandbox with the app, its database and the browsers running, snapshot it, and run one test shard in each copy.
- How to run Jupyter notebooks headless in the cloudUpload the notebook to a sandbox, run it with papermill or jupyter nbconvert --execute, and download the executed copy.
- How to run linters and security scanners on your own code with an AI agentRun Ruff, Semgrep and npm audit on your code in a sandbox, give the model the JSON findings, and rescan to confirm each fix.
- How to run LLM-generated transformation code in a data pipelineRun each batch's model-written transform in its own sandbox with the internet off, check the output's shape, and only then load it.
- How to run scientific computing jobs in cloud sandboxesBuild an image with your numerical stack, give each simulation or parameter its own sandbox, and read the results back as files.
- How to run SWE-bench-style agent evals in parallel sandboxesBuild each task's environment once, start one isolated sandbox per attempt, apply the agent's patch, run the tests and record pass or fail.
- How to run tree search over an agent's environment with sandbox forksMake each sandbox a search node: fork it per candidate action, score the children, pause the best and stop the rest.
- How to run untrusted code from an LLM safelyRun it in a throwaway microVM with its own kernel, no route to your network, a time limit and a spending cap.
- How to run visual regression tests with stable screenshotsMake and compare screenshot baselines in the same pinned sandbox image, so the OS, fonts and browser never change between runs.
- How to train small machine learning models on CPU sandboxesBuild an image with scikit-learn or CPU-only PyTorch, train in a sandbox with every vCPU busy, and copy the model file out at the end.
- Sandbox for a coding agent: give an AI agent its own computerGive the agent a Linux microVM with the repository cloned, let it run commands and tests there, and take back the diff when it is done.
- Sandboxes for RL environments: run agent rollouts in parallelPrepare the environment once as a snapshot, start a fresh copy for each episode, let the policy act, score the end state, and stop the copy.