Context

Any agent with write access to a real machine. The moment it can run arbitrary commands, the question stops being "will it make a mistake" and becomes "how much does the mistake cost".

Problem

Autonomy and safety pull in opposite directions. Supervise every command and you have a very slow autocomplete; grant unrestricted shell and you cannot leave the room. Most teams oscillate between the two and settle on whatever their tolerance for anxiety allows.

Pattern

Decide the blast radius before the run, in the same way you would for a deploy:

  • Filesystem scope - the workspace, nothing above it. No home directory, no credentials, no package manager writing outside.
  • Network - off by default. Turned on per task when the task genuinely needs it.
  • Destructive operations - delete, force-push, mass-rename: either denied, or gated behind explicit approval with the exact command shown.
  • Reversibility - a checkpoint before the run: a commit, a branch, a snapshot. The test is whether "undo everything this run did" is one command.

Then size the task to the radius: if a mistake would touch one module, let it run unattended. If a mistake would rewrite the build system, watch it.

Tradeoffs

Every restriction is a wall you will hit at 11pm on a task that genuinely needed network access. Granular permissions add configuration burden and a new class of confusing failures - the agent spends three turns diagnosing what is actually a denial.

When not to use it

Read-only work - analysis, review, research - where there is nothing to break. Also throwaway environments, where the cheapest policy is to make the environment disposable rather than to police it.