Skip to slide
072 / 170

CH 07 · Sandboxing, Approvals, and Checkpoints · 03/06

Read as one page

How it works: a policy engine in front of execution

The cleanest design puts a single decision point in front of every tool call. Recall from Chapter 6 that ToolRegistry.call had a marked spot for "step 3: permissions." Here we build the ApprovalPolicy that goes there. It classifies a command into a tier and returns allow, deny, or ask. For unattended runs, "ask" can be delegated to a guardian function instead of a human.

A policy engine in front of execution
A policy engine in front of execution

Field check

See it worked. Then try it.

Evaluate the actual requested operation at the execution boundary, even if the prompt already describes a policy.

Worked examples

A tool asks to read a file inside the approved workspace.

The policy engine can allow that specific operation within its scope.

The next tool call asks to delete a protected directory.

Evaluate this new operation separately; earlier permission to read does not authorize a destructive write.

Your turnA model emits a shell command outside the allowed filesystem scope. Where should the permission decision be enforced?

Finished this idea?

Your learning progress stays in this browser.