THE SHORT ANSWER
Capability is not authorization; agents can execute destructive actions if guardrails don't match permissions.
Flashcards
Q1
What is context degradation in long agent loops?
The loss of original intent as the agent fills its context window with intermediate tool outputs, leading to hallucinations.
Q2
How do you secure tool execution environments?
By applying the principle of least privilege, using sandboxes, and requiring human approval for state-mutating actions.
Q3
Why is human-in-the-loop (HITL) mandatory for state-changing workflows?
Because AI models lack real-world accountability and cannot assess the business impact of deleting data or sending emails.
Related Concepts
Chaos Stack Field Notes FAQs
What are Chaos Stack Field Notes?
Chaos Stack Field Notes are technical flashcards that explain core engineering concepts quickly.
How are these different from topics?
Topics are broad thematic hubs that connect characters, episodes, and environments. Field Notes are short, direct Q&A flashcards for quick technical alignment.
AI Summary
This page covers AI Workflow Autonomy Risks as a technical flashcard. Description: What is the primary risk of autonomous agent tool-calling?.