Docker has published the Sandbox Kit Specification v3, a format that turns the configuration of an AI agent's runtime environment into a portable, versioned artifact. Instead of hand-assembling a Dockerfile plus loose config for every agent, the Kit spec lets you declare the agent's network rules, credentials, and volumes together and distribute the result like any other container image.
The core idea is reuse: a Kit is described as an ordinary, pinnable OCI image. Because it conforms to OCI conventions, it can be stored in an existing registry, pulled by digest, and referenced from a Dockerfile-style workflow without a bespoke delivery mechanism. Key concrete points from the announcement include:
- Sandbox Kit Specification v3 is the current version of the format for defining an AI agent's sandbox.
- A Kit bundles network rules, credentials, and volumes into a single declarative unit.
- Kits are packaged as OCI images, so they are pinnable by digest for reproducible environments.
- The move extends the familiar
Dockerfile-to-image model to agent sandboxes, rather than requiring custom tooling.
For developers, this addresses a recurring pain point in agent work: environments that are hard to reproduce or audit. Network rules and credentials are exactly the parts you want reviewed, pinned, and rolled back deliberately, not re-created ad hoc on each run. Treating them as an OCI artifact means existing registry scanning, signing, and promotion workflows can apply. It also lowers the barrier to sharing a known-good agent sandbox across a team, since the artifact travels through channels teams already operate. The spec reinforces Docker's broader push to make agent infrastructure as boring and reproducible as ordinary container images.
Source: https://www.docker.com/blog/docker-sandbox-kit-spec/