LockedIn FDE Training field practice · Free and open

Separate a proposal from permission

Follow a model's recommendation to the boundary that permits an action.

Read the concept. Explore two decisions. Leave with a test you would run next.

The working model

Understand the boundary.

An agent can propose a tool call; your application decides whether to execute it. For a tool your application runs, enforce the user's authority and the operation's limits at execution. A fluent explanation cannot grant permission.

Propose
Model requests an operation
Authorize
Server checks scope and approval
Execute
Approved operation gets a receipt

A worked example

A fictional support assistant may draft a reply. Its server receives a proposed send_reply call, checks the caller, the request and a human approval tied to that exact draft, then permits or refuses the send. Editing the draft invalidates the earlier approval.

Claude: how application tools execute

Primary reading checked 2026-10-10. Provider learning stays with its issuer.

Practice deskDecision 1 of 2

Make a decision

The assistant produces a useful reply and requests send_reply. The user has approved reading the support request, but has not approved sending this draft.

What should the application do?

Choose one response before comparing the reasoning.Hold the send and request approval of this exact draft.Send if the model is confident in the reply.Tell the model to ask permission, then execute every tool call it returns.

Another concept

Prove that a repair fixes the defect

Open the next free exercise

For teams

Build evidence through real work.

Advanced lessons, labs, mentoring and qualification workflows sit behind provisioned internal or licensed seats. Public practice does not unlock them.

Explore the private training pathway