Exercise 6: Sandbox & Governance¶

Duration: 15 min | Module: 4 — Multi-Agent & Advanced

Objective¶

Run agy in --sandbox mode for safe code audits, understand the --dangerously-skip-permissions flag, and model a governance-appropriate workflow for enterprise environments.

Part 1: Sandbox Mode — Safe Audit (7 min)¶

Run a security audit with terminal restrictions enabled:

agy --sandbox \
    --print "Scan this entire codebase for: (1) hardcoded secrets or API keys, (2) SQL injection risks, (3) insecure direct object references, (4) any .env files or credentials committed to the repo. Output findings as markdown with severity levels." \
    --print-timeout 5m > audit-sandbox.md

cat audit-sandbox.md

Key properties of this run:

--sandbox restricts terminal command execution — agy reads files but cannot run arbitrary shell commands
--print means no interactive session — purely automated
Output is captured to a file for audit trail

When to use this pattern:

Auditing code you don't fully trust
Compliance scanning in regulated environments
Running on production codebases where side effects are unacceptable

Part 2: Auto-Approve Mode — Understand the Risk (5 min)¶

--dangerously-skip-permissions bypasses all tool approval prompts. agy executes file writes and shell commands without asking.

Safe demonstration: run it with --sandbox to show auto-approval without actual command execution:

agy --sandbox --dangerously-skip-permissions \
    --print "List all TODO comments in this codebase and generate a prioritized backlog." \
    --print-timeout 3m

Without --sandbox, this flag would allow agy to write files, run tests, and execute commands without prompting. Only use it:

In CI/CD where no human is present
Paired with --sandbox for read-only audits
On throwaway environments where writes are acceptable

Never in production

--dangerously-skip-permissions without --sandbox in an interactive session on a live codebase is a footgun. There's no undo for overwritten files.

Part 3: Governance Workflow (3 min)¶

Model a two-phase governance workflow:

Phase 1: Safe analysis (no side effects)¶

agy --sandbox \
    --print "Analyze all database operations in this codebase. Flag any that lack transaction safety or input validation." \
    --print-timeout 3m > phase1-analysis.md

Phase 2: Human reviews, then approves interactive session¶

cat phase1-analysis.md  # human reviews findings

# If approved, continue with interactive session for remediation
agy -i "Based on the findings in phase1-analysis.md, fix the top 3 database safety issues."

This pattern is the enterprise-grade model: read without trust, write only after review.

Completion Criteria¶

[ ] agy --sandbox --print "..." ran and produced an audit file
[ ] Understand when --dangerously-skip-permissions is appropriate vs dangerous
[ ] Implemented the two-phase governance workflow (sandbox audit → human review → interactive fix)