Homework — Design a Sandbox for a Ticket-to-PR Agent
~25 minutes · sandbox design exercise
This homework is a design exercise, not a coding exercise. You will write a sandbox specification for a realistic autonomous agent and then check it against the reference answer. The point is to force yourself to make every policy explicit before you look at how someone else answered it.
Do not open answer.html until you have written your own five policies. The worked reference is there to check your reasoning, not to seed it.
Scenario & design questions
You are designing an autonomous AI agent that reads issues from a ticket system and opens PRs to fix them. The agent runs the model, git, and a test runner. Design its sandbox. Specify: (1) network policy — what it must reach and what it must not; (2) filesystem policy — what's read-only, what's read-write, where secrets live; (3) capability policy — what caps, what user; (4) secret scoping — what tokens, what scopes, how rotated; (5) escape budget — what's the worst case if the agent is fully prompt-injected?
Write out each of the five policies in one or two paragraphs. Be specific: name the paths, name the caps, name the token scopes. "It should be locked down" is not an answer.
✓ Self-check: you have written all five policies, every policy names concrete mechanisms (not aspirations), and your escape-budget answer describes the worst-case outcome under the assumption that the model is fully compromised — not the best-case outcome under the assumption that it isn't.