Skip to content
Agentguiden.no
Submit

Sandbox your coding agents’ tests with Claude Code

Marut Pandya had 7 Claude Code agents working in parallel with no sandbox. On 28 October his Mac started reading JSON aloud.

Done with Claude CodeModerate

What was done

Marut Pandya had 7 Claude Code agents working in parallel with no sandbox. On 28 October his Mac started reading JSON aloud. One agent had written a test with a one-line bug, and it started 4 real processes. Two spoke, and two were meant to upload data to his production server. They sent nothing, but the agent fixed the bug in 7 seconds without mentioning the processes.

How to do it

  1. Paste the prompt into the tool, and replace the text in brackets with your own.
  2. Read the answer through. Do not send, delete or publish anything until you have seen it yourself.
  3. Check the original before you trust a figure, a name or a date.

Try it with Claude Code

Before you run tests in [project], check that you are in [sandbox or container] that can only write to the project folder. Use test addresses, never [production address], and no real keys. If a test starts real processes, sends anything over the network or touches files outside the folder, stop and tell me. List every unexpected side effect in your report. You are done when tests pass and the report lists side effects.

Replace text in brackets with your own.

Done when. Tests pass and the report lists side effects.

Read the original

Marut Pandya ·

Five agents a day

A short email each morning with five concrete examples of what people use AI agents for. Each example links to its source. You can unsubscribe at any time.

You will first get an email to confirm the address.