Tools give data and actions
Tools do two jobs: fetch data the model doesn't have (an order status) and take actions it can't take (create a Jira ticket).
Last updated
Analogy
GET routes and POST routes. Reading is cheap and safe; writing deserves more care.
In an agent
Read tools can run freely. Write tools (delete, refund, send email) usually need validation and a human approval step.
Questions people ask
What kinds of tools does an agent need?
Two kinds. Read tools fetch data the model does not have, like an order status or a document search. Write tools take actions it cannot take, like creating a ticket or sending an email. Read tools are low risk. Write tools change the world, so they need more checks.
Should write tools need human approval?
Usually, yes, for anything that is hard to undo or reaches customers: deleting data, refunds, payments and outgoing emails. The agent proposes the action and a person approves it. Read tools can run freely. This keeps the damage small if the agent is confused or tricked by injected text.