Agent Brief

Simple Definition

An Agent Brief is a structured specification that defines an AI agent’s goal, trigger, inputs, steps, permissions, stop rules, success criteria, and test cases, written down before the agent is connected to any real tools.

A Simple Structure

Goal: what the agent is actually trying to accomplish
Trigger: what starts the agent’s work
Inputs: what information or context it receives
Steps: the general approach it should follow
May do automatically: actions that don’t need approval
Must ask before: actions that require a human check first
Must never do: hard limits, regardless of instructions it encounters
Stop rules: when the agent should stop, on success, on failure, or on a budget limit
Success check: how anyone, including the agent, can tell the work is actually done
Output format: what a finished result looks like
Test cases: example inputs and expected behavior to sanity-check the design

Why It Matters

Writing an Agent Brief before connecting an agent to tools forces the job to be defined before the agent receives power. It’s much easier to notice a missing constraint or an overly broad permission on paper than after an agent has already acted on it. A brief also gives you something concrete to test against, instead of judging the agent by vibes after the fact.

Example

A brief for an “inbox triage” agent might specify: goal (flag urgent emails and draft short replies to routine ones), trigger (new email arrives), may do automatically (label and sort), must ask before (sending any drafted reply), must never do (delete emails, forward anything externally), stop rule (after processing the current inbox, not continuous polling), and success check (a person reviews the labels and drafts once a day and finds them accurate).