Container-Use

Agent sandboxing primitive by Dagger. Each agent gets a containerized environment that physically cannot reach the host system. Founded by Solomon Hykes — the creator of Docker — which is the source’s “pay attention” framing for the project.

What it does

  • Each agent runs in its own isolated container
  • Filesystem, network, and process scope are scoped to the container
  • An agent that goes rogue and starts deleting files can only delete files inside its own sandbox

Why it matters in the wiki (load-bearing for hardening)

This is the wiki’s first concrete tooling answer to rule 1 of ai-personal-agent-hardening: human-in-loop quarantine is mandatory. Container-Use turns “the agent must run in a quarantine” from a principle into an installable primitive. Combined with promptfoo (prompt-level red-teaming), the wiki now has both ends of the defensive stack from a single source:

  • Promptfoo = test the prompts before the agent ships
  • Container-Use = contain the agent after it ships

Sits adjacent to the openclaw / managed-agents / conway thread on what “always-on autonomous agent” infrastructure actually requires: a sandbox the agent cannot escape is one of the load-bearing primitives.

Type and platform

  • Type: CLI / runtime
  • Platforms: model-agnostic; runs as host-side infrastructure under any agent
  • Status: tracked

Hands-on notes

(empty)

Sources

See Also