AI engineering glossary · Agents tools MCP
What is guardrails for agents?
Guardrails for agents: Checks and limits placed around an agent to keep its behaviour safe and on task, such as input filters, output validators, allowed-tool lists, spending caps and step limits. They reduce harm when the model errs or is manipulated.
Guardrails for agents: an example
A guardrail blocks any tool call that would transfer more than Rs 10,000 and refuses requests unrelated to banking.
Related terms
More Agents tools MCP terms
- Human-in-the-loop
- Human-on-the-loop
- Idempotency (tool calls)
- Indirect prompt injection
- Interrupt (agent)
- JSON-RPC
- Kill switch
- LangChain
Where you learn it
Guardrails for agents is part of the AI Engineering Program, which teaches it with notes, a hands-on lab and an exam. The first lesson is free.