I wanted AI agents to work for me in the background like a coworker, not require constant supervision. TSK lets you queue tasks for Claude Code or Codex in sandboxed containers - when they finish, you get a git branch to review. This frees up your time to focus on more interesting problems without the constant interruption of agent approvals.

As an example, here's what it looks like to kick off a new task:

    tsk run --type feat \
        --name greeting \
        --description "Add a greeting for users each time they run a command in TSK"

For this command, TSK will make a copy of your repo, create an isolated docker container with the tech stack you are using, start a forward proxy to limit internet access, mount your Claude Code configuration and repo into the container, and finally kick off the coding agent. Internet access is limited to agent API endpoints and common tools for managing dependencies e.g. crates.io or pypi.

The "feat" task template includes repetitive but important instructions like "write unit tests", "update documentation", and "write a detailed commit message". The "description" flag gets injected into the template. It's easy to add new task templates. They are very similar to slash commands in Codex or Claude.

When the agent finishes, a new commit and branch is non-disruptively created in your repository for you to review.

Additionally, you can also use the `tsk server` to run multiple agents in parallel. This command will tell the TSK server to run Claude Code and Codex in parallel with the exact same instructions:

    tsk add --type refactor \
        --name clippy \
        --description "Fix all the clippy linter errors in the codebase." \
        --agent claude,codex

TSK also supports a `tsk shell` command that starts up an interactive shell in one of the sandboxes. This works great with a multiplexer to manage multiple interactive agents.

I just finished Codex support and multi-agent comparisons today so thought it was a good time to share!