Permission request

A permission request is the moment an agent stops and asks before it does something that matters. It has decided to make a tool call, run a shell command, or overwrite a file, and instead of just doing it, it shows you what it is about to do and waits for a yes or no.

Why the pause matters

An agent that acted on every impulse would be alarming. One wrong command can delete files, drop a table, or push broken code. The permission request is the point where a person gets to catch the mistake before it happens. It is the concrete way an agent keeps a human in the loop.

What usually triggers one:

• Running a command that changes the system: installs, deletes, network calls.
• Writing or overwriting files outside a safe scratch area.
• Anything the tool's policy has flagged as too risky to auto-approve.

Read them, do not rubber-stamp

The whole value of a permission request disappears if you approve on autopilot. Read what the command actually does. Notice the rm -rf, the unfamiliar URL, the write to the wrong path. The prompt only protects you if you are actually looking.

The flip side is real too: too many prompts and you stop reading them. That tension is exactly what the permission mode exists to manage.

Where it lives in the loop

The harness decides whether a requested action needs a human yes before it runs. That pause is the permission request.

async function runToolCall(block) {
  if (isRisky(block.name, block.input)) {
    const ok = await askUser('Allow ' + block.name + '?') // the permission request
    if (!ok) {
      return { type: 'tool_result', tool_use_id: block.id, content: 'Denied by the user.' }
    }
  }
  return execute(block)
}