# The Agent Constitution

**One page. Copy it, fill in the blanks, and put your name on it.**

---

## Purpose

This document authorizes one named person to run one named agent inside stated limits,
and it prevents that agent from contacting a customer, changing a system of record, or
acting on information it has not verified, unless a person has approved that specific
action first.

---

## The permission ladder

Every agent starts at rung 1. An agent moves up only on recorded evidence. Rung 3 is a
ceiling, not a step. Rule 4 applies at every rung.

### 1. Draft only, until proven

The agent produces output. A person reviews it, and that person is the one who sends,
saves, or acts. The agent cannot complete an action by itself.

Every new agent starts here. There is no fast track and no exception for an agent that
worked somewhere else.

**Evidence to move up:** ______ consecutive runs with no correction needed, reviewed by
______________________, with each review recorded and dated.

### 2. Promotion on evidence

An agent may be promoted to act on **internal** systems without a person pressing the
button each time. Internal means a CRM field, a task queue, an internal dashboard, or a
channel only employees can see.

Promotion is a decision someone makes and signs, based on the runs recorded at rung 1.
It is not automatic and it is not a default that arrives with a new tool version. An
agent that has been changed goes back to rung 1.

**Evidence to move up:** not applicable. Rung 2 is as far as an agent goes.

### 3. Outbound is always human

Anything that reaches a customer is approved by a named person before it goes. Email,
in-app message, review request, community post, calendar invite, anything published
under the company name.

No number of clean runs promotes an agent past this rule. It does not expire, and it
does not get waived for volume, for a deadline, or for a campaign.

**Evidence to move up:** none exists. This is the ceiling.

### 4. Plain-file context

The agent reads its instructions and its context from one canonical, human-readable
file that a person can open, read, and edit without a developer.

No hidden prompts, no context living in a chat history, no rules that exist only inside
a vendor's interface. If a reviewer cannot read what the agent was told, the agent is
not reviewable, and an agent that is not reviewable does not run.

**Applies at:** every rung, permanently.

---

## What the agent may never do

Regardless of rung, and regardless of who asks:

- Send an external communication without a person approving that specific message.
- Act on data it has not verified against the source.
- Create, change, or delete a record in a system of record without approval.
- Represent itself as a person to a customer, or sign as one.
- Move money, agree to a commitment, or accept terms.

---

## Verification

The agent checks its output against the real source, and where it cannot confirm
something it says so and stops. It does not fill a gap with a reasonable guess.

The failure this prevents is specific. An agent asked to thank event attendees read the
invite list rather than the check-in list, and drafted warm thank-you notes to people
who had not turned up. Every sentence was well written. The underlying fact was wrong,
and no amount of review of the writing would have caught it, because the writing was
not the problem.

An agent that flags what it could not confirm is doing its job. An agent that never
flags anything is not being careful, it is being confident.

---

## Named accountability

Every agent has one human owner, by name, not a team and not a role. The owner is
accountable for what the agent produces, reviews it on a stated cadence, and is the
person who answers when someone asks why it did what it did.

An agent whose owner leaves, changes role, or stops reviewing is switched off until a
new owner takes it on. An agent without an owner does not run.

---

## Adoption block

Fill this in per agent. One page per agent, kept where the team can find it.

| Field | Entry |
|---|---|
| Agent name | |
| What it produces | |
| Human owner (name) | |
| Current rung | 1 / 2 |
| Data sources it reads | |
| Canonical context file (location) | |
| Promotion threshold | ______ clean runs, reviewed by ______________ |
| Date promoted to current rung | |
| Review cadence | |
| Next review date | |
| Approved by | |

---

*Adapted from the Customer-Led Growth agent framework. Kevin Lau, The Customer Continuum.
Use it, change it, put your own name on it.*