Agent Safety
Domain: Agent Architecture
Motivation
Agent safety is the set of guardrails around agent behavior: sandbox and approval policies that gate operations, the Auto review layer that assesses each tool call, and the fixed permission scope of a delegated subagent that cannot be widened from inside its session. These bounds are what keep a fully-permitted agent from overstepping.
Concrete Example
A delegated subagent’s runtime-context snapshot carries the delegation-scope statement: its permission scope is fixed at start, and operations needing approval are rejected automatically.
Analogy
Handcuffs, checkpoints, and a supervisor — the controls that keep a worker inside the line.