Module 15 of 16 · 18 min
Agents, Tools, and Guardrails
Understand what changes when a model can take actions.
Core concept
By the end
You will be able to
- Distinguish a chat response from an agentic loop.
- Identify tool, permission, and verification boundaries.
- Recognize when human approval is required.
01
An agent observes, decides, acts, and checks
An agent combines a model with tools and a loop. It can inspect state, choose an action, use a tool, evaluate the result, and continue toward a goal.
Capability increases risk. Read access, file edits, messages, deployments, and destructive operations should have different permission and approval policies.
02
Guardrails belong around the model
Reliable agent systems constrain tools, scope credentials, validate inputs and outputs, record actions, and stop for human decisions when consequences are difficult to reverse.
Evidence
Sources and verification
- Building effective agentsAnthropic · verified 2026-07-23
Knowledge check
Make it stick.
Choose the strongest answer for each question. Your attempts become part of your device-local transcript.