Topic 6
People in the loop
So far the agent has generated, validated and executed its own plans. AI Engineering (Ch. 6, "Planning overview") points out that people can be involved at any of those stages, both to help the agent and to manage risk.
Predict first
Scroll sideways to see the whole diagram
AI Engineering: humans can be involved at any stage. For hard tasks, an expert can give the high-level plan for the agent to expand.
Text version
- AI Engineering: humans can be involved at any stage. For hard tasks, an expert can give the high-level plan for the agent to expand.
- A person can review a plan before it runs.
- For risky operations, such as updating a database or merging code, the system can ask for explicit approval, or leave the step to a person.
- To make this work, the book says, clearly define the level of automation an agent can have for each action. Lesson 6 builds that table.
The book describes a role for people at each stage:
- Planning: for complex tasks the agent struggles to plan, a human expert provides a high-level plan and the agent fills it in.
- Validation: a person reviews the plan before execution.
- Execution: for risky operations, such as updating a database or merging a code change, the system asks for explicit human approval, or lets a person carry out the step.
A checkpoint is a harness feature, not a prompt. The harness pauses the loop, shows a person the proposed action with its arguments, and continues only on approval. That means the agent's state must survive the wait, possibly for hours, which is lesson 4's checkpointing.
Quick check