Editing Agents vs. Assistants (section)

== The key question: how much autonomy should you give? ==

More autonomy isn't always better. The right level depends on three things:

'''Stakes.''' How bad is it if the AI gets it wrong? For brainstorming ideas, high autonomy is fine — a wrong suggestion costs nothing. For sending an email to a client, you want to review before it sends. For financial calculations, you verify every number.

'''Predictability.''' How well-defined is the task? Formatting a weekly report from the same data sources is highly predictable — good candidate for an agent. "Help me figure out our strategy for next quarter" requires judgment at every step — keep it as a collaborative conversation.

'''Your expertise.''' Can you evaluate the output? If you're an expert in the domain, you can give AI more autonomy because you'll catch mistakes quickly. If you're learning a new area, keep the AI in assistant mode where you're directing every step and building your own understanding.

A practical framework:

| Autonomy level || Use when || Watch out for

| '''You direct, AI executes''' || High stakes, new domains, learning || Slower, but you understand everything
| '''AI proposes, you approve''' || Medium stakes, familiar territory || Review carefully — don't rubber-stamp
| '''AI acts, you spot-check''' || Low stakes, predictable tasks, repeatable workflows || Set up verification checkpoints
| '''AI acts autonomously''' || Very low stakes, highly predictable, easily reversible || Only if you can undo mistakes cheaply