Keep a Human in the Loop

Agent autonomy is not binary. It is a dial you can turn:

1. Suggest only. The agent proposes. You decide.
2. Approve required. The agent shows its plan. You approve each step.
3. Auto within rules. The agent acts within a bounded scope.
4. Fully autonomous. The agent runs and reports.

Most teams should stay at 1 or 2 for a long time.

Agents are good at pattern matching and fast execution. They are bad at:

• Judgment under uncertainty
• Understanding business context
• Evaluating tradeoffs that are not encoded
• Recovering from novel failures
• Knowing when not to act

A human in the loop catches the cases the agent does not know it cannot handle.

Start with low-stakes, reversible actions:

• Formatting code
• Generating summaries
• Running test suites
• Drafting messages
• Updating internal documentation

Only move to high-stakes actions after consistent success:

• Committing to main
• Sending external email
• Modifying production config
• Spending money or using credits

Good approval flows show:

• What the agent is about to do
• Why it thinks this is the right action
• What could go wrong
• How to undo it
• A clear approve / reject / modify path

Bad approval flows show a wall of JSON and a yes/no button.