Vocabulary for building AI agents that plan, act, and recover from failure in production systems.
An isolated execution environment where an agent's actions, like running generated code, are contained and can't touch the real system or network.
Limiting exactly which actions and data an agent's tools are allowed to touch, so a bug or bad decision can only cause damage within that narrow boundary.