Human-in-the-loop puts a person at decision points inside the agent's flow.
Approval gates should sit before anything irreversible or externally visible.
Sending email, spending money and deleting data are the classic gates.
The agent pauses, presents its intent, and waits for confirmation.
It is the most effective safety control available for agents that act in the real world.
Human-in-the-loop puts a person at chosen decision points inside an agent's flow.
It is the most effective safety control available for agents that act in the world.
The rule is simple: gate anything irreversible or externally visible.
Sending an email, spending money, deleting data, changing production configuration.
The agent pauses, states what it intends to do, and waits.
Showing the exact action rather than a summary is what makes the review meaningful.
About to send this email, to this address, with this text.
A vague description invites approval without reading, which is worse than no gate.
Approval fatigue is the real enemy, since a person clicking yes fifty times stops reading.
Gate few things and gate them well.
Auto-approve the reversible and the low-value, and escalate the rest.
Thresholds work well, such as refunds under twenty pounds proceeding automatically.
Provide an undo where you can, since reversibility is worth more than a confirmation.
Log every approval with who approved it and what they saw.
As agents get more capable this becomes more important rather than less.
One question on this card. Every option explains itself once you pick, right or wrong.
17 concepts in this module, then a graded test with every answer explained.