Lesson 03 // 9 Minutes

When a Model Becomes an Agent

An agent is a model inside a loop with tools, state, and a goal. Each added capability also adds a new failure path.

Four Parts of an Agent

The model proposes a next action. Tools let it read, calculate, browse, write, or execute. Memory carries selected state forward. A control loop evaluates what happened and decides whether to continue.

None of these parts guarantees good judgment. A mistaken plan can become a mistaken action faster.

Capability Should Follow Permission

Give each tool the least access required. Reading a folder does not require permission to delete it. Drafting an email does not require permission to send it. Planning a deployment does not require production credentials.

Control Rule

Separate reversible preparation from consequential execution. Require explicit approval before sending, spending, deleting, publishing, or changing access.

Bound the Loop

Set time, cost, action, and retry limits. Log actions without storing unnecessary private content. Stop on ambiguity instead of improvising around a failed safety check.

Five-Minute Exercise

Pick one task you might delegate. List every tool it needs, the minimum permission for each, the maximum number of steps, and the actions that require your approval.

Next: Responsible AI