Part

Runner

You might say: worker, execution environment, build machine, server, box, sandbox

The disposable machine or container that executes one run. It supplies the filesystem, tools, network path, and credentials that the task can reach.

What it does

Turns a queued run into an isolated working copy and returns its files, status, and evidence.

Ask

What can this run reach, and what survives when it ends?

Typical

One runner per run is the safe default. Choose a CI worker, a sandbox API, a bundled agent platform, a worker you operate, your own container cluster, or a local worktree for a few runs on one machine.

Trouble

Two runs share files or credentials, a worker stays alive after failure, or a task depends on state from an earlier run.

What to look for

1. The runner image, machine size, and status appear on the run page
2. A fresh workspace path and a run timer
3. Runner labels or a sandbox provider on the run page

Parts

• Disposable environment: A fresh environment is created for one run and then wiped or discarded.
• Run timeout: The maximum time a run or one of its steps may execute before the runner cancels it.
• Cache: An optional store for dependencies or intermediate files that a run can regenerate.
• Network egress: The destinations and protocols a runner may reach outside its boundary.
• Secrets: Credentials made available to a run through a controlled scope.
• Security boundary: The set of gates around a run that separate trusted factory services from task input, tools, data, and side effects.

Catalogue: Runner type

• CI runner: A hosted CI service runs each job on a fresh instance of a runner image. The CI provider hosts the worker and starts a new one for every job.
• MicroVM sandbox API: An API creates a strong isolation boundary for code, then destroys or pauses the sandbox by policy. The service exposes a sandbox boundary as an API call.
• Bundled agent platform: The agent service provisions a Linux workspace, runs the task, and returns its artifacts through one product surface. Agent execution and environment provisioning arrive as one managed service.
• Bring-your-own worker: You register a machine you operate as a self-hosted runner and choose its labels, image, and cleanup. The operator supplies the host while the orchestrator still assigns jobs.
• Own container cluster: You run ephemeral runners on a container cluster you operate. A controller such as Actions Runner Controller creates a runner for each job and cleans it up afterwards. The operator owns the cluster that hosts the runners, and automation gives each job a clean environment.
• Local worktree: The run works in its own git worktree on a developer's machine, with its own files and branch. Files are isolated, while the machine, its ports and its credentials stay shared, so a laptop holds only a few runs at once. The worktree isolates the working copy, and nothing else.

Sources

1. GitHub, Choosing the runner for a job, Choosing GitHub-hosted runnersIf you use a GitHub-hosted runner, each job runs in a fresh instance of a runner image specified by `runs-on`.
2. GitHub, Self-hosted runners reference, Ephemeral runners for autoscalingWith ephemeral runners, this can be guaranteed because GitHub only assigns one job to a runner.
3. E2B, Security and compliance, Isolation and tenancySandboxes are destroyed on timeout or shutdown. Filesystem and memory state can be preserved across a pause and restored on resume.
4. Daytona, Isolation, IsolationDaytona sandboxes are isolated by default. Code running in a sandbox cannot read another sandbox’s filesystem or memory, is not on a shared network with other sandboxes, and its credentials and API access are scoped to its own organization.
5. Modal, Sandboxes, SandboxesSandboxes have a default maximum lifetime of 5 minutes. You can change this by passing a `timeout` of up to 24 hours to the `Sandbox.create(...)` function.
6. Vercel, Vercel Sandbox, Run agent-generated code in isolationRun untrusted or agent-generated code in isolated Linux microVMs. Use Vercel Sandbox for agent workflows, debugging, and one-off commands.
7. Matt Pocock, 11 Tips For AI Coding With Ralph Wiggum, 9. Use Docker SandboxesThis runs Claude Code inside a container. Your current directory is mounted, but nothing else.

Use it on your project

teardown-apply: hold "Runner" (runner) from the teardown at https://teardownatlas.com/software-factory/ against my project

Paste it into a coding agent that has the teardown skills.

An AI agent wrote this map with the teardown skill. It can contain errors: check the sources before you rely on a part. About and legal