Tinker Agent sandboxing

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

...

Nothing run yet.the code runs in the sandbox · files: working copy writable, data read-only, keys hidden · network: allow-list through the proxy · secret: in the sandbox

Chapter 6 of 7

Fewer prompts, same safety

Before sandboxing, a careful team asked a person to approve almost every action the agent took. That is slow, and people wave things through. The sandbox changes the maths: the boundary does the enforcing.

This chapter: With a sandbox boundary, which actions still need a person, and which can run freely?

You will learn

  • Why a sandbox lets most actions run without a prompt
  • That a person is asked only when an action would cross the boundary
  • Why approving a boundary crossing hands the safety back to a tired human
Words you’ll meet (3)
Approval
A person who says yes or no to an action before it runs.
Boundary crossing
An action the sandbox would stop, such as a fetch to a host not on the allow-list.
Prompt fatigue
When a person approves so many actions that they stop reading them.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 5

Before sandboxing, a careful team approved almost every action. Now the agent runs an ordinary report task inside the sandbox, with no approvals turned on. Run it and notice: every action stayed inside the boundary, so nobody was asked anything.