liveBook · Manning

Least agency

As explained in Designing AI Agents

A principle of granting an agent only the autonomy needed to perform its delegated task.

Least agency grants an agent only the autonomy required for its delegated task. It limits how much harm a wrong decision can cause, rather than relying on the agent to choose correctly every time. section 9.1

Limit compensation authority

An agent may be allowed to issue one 100-yuan coupon without authority to change the compensation rule for every customer. A per-case ceiling, daily budget, retry limit, and cancellation or compensation path help contain a bad decision. section 9.1

A gate that checks each tool call separately can miss a harmful objective assembled from individually acceptable tasks. Validate user and agent intent at runtime and preserve that intent across the sequence, within the outer boundary set by least agency. section 9.1

How Designing AI Agents covers it

How deep it goes
Explains an autonomy boundary through operational limits and sequence-level intent validation.
Examples
Illustrates bounded authority with a 100-yuan coupon and contrasts it with changing customer-wide compensation rules.
What you can do after
Evaluate agent autonomy against delegated tasks and identify limits that contain bad decisions.
How it connects
Connects agent autonomy with agent goal hijacking, runtime intent validation, and tool-call gates.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.