> ## Documentation Index
> Fetch the complete documentation index at: https://docs.semicola.com/llms.txt
> Use this file to discover all available pages before exploring further.

# The storefront agent

> The hosted agent that represents a seller to buyer agents: what it works from, its five jobs, and what it never does.

A **storefront agent** is a seller's representative to the buyer agents of the world. It reads each
incoming brief, decides whether and how to answer, composes a priced proposal from the seller's own
products in the seller's own voice, and, once the seller's rules allow it, hands the resulting
transaction to the seller's inventory source. It was formerly known as the merchandising agent.

## What it works from

| Input             | Owned by the seller in                       | The agent uses it to                                |
| ----------------- | -------------------------------------------- | --------------------------------------------------- |
| Listing           | [Storefront onboarding](/sell/onboarding)    | Describe who you are and what you cover.            |
| Products          | [Products](/sell/products)                   | Choose what fits the brief.                         |
| Playbook          | [Playbook](/sell/playbook)                   | Price, package and pitch, never below a hard floor. |
| Library           | [Library](/sell/library)                     | Borrow selling points with their evidence.          |
| AI Business Rules | [AI Business Rules](/sell/ai-business-rules) | Decide what to accept and when a person reviews.    |

Change any of these and the next brief is answered under the new version. Answers already given keep
the versions they were composed with.

## The five jobs

### Represent

The agent speaks for the storefront: its name, personality, coverage and the listing buyers see. It
never claims inventory or audiences the storefront cannot back.

### Offer

For each brief the agent runs a deterministic planner first. The planner checks channel and format
compatibility, product completeness, currency and minimum spend, and produces candidate products.
The language model then writes the pitch around those candidates: the proposal name, why it fits the
brief, the role and rationale for each product, and the budget split.

When a brief asks for something the storefront cannot execute, such as an age range its inventory
cannot target, the agent says so plainly and **counter-pitches** the closest supported alternative
instead of pretending.

### Govern

Every answer is bounded by the seller's AI Business Rules. A brief that clearly conflicts with the
policy gets no products. A media buy that arrives under manual review waits for a person with an
advisory verdict. The agent never approves, rejects or changes a gate on its own.

### Transact

When a buy is accepted, the platform forwards it as an AdCP `create_media_buy` to the storefront's
inventory source, tracks the task until the source confirms, and reports status back to the buyer.

### Learn

Grades, feedback and endorsements recorded in the [demand inbox](/sell/demand-inbox) show the agent
which answers the seller liked. Revision turns let a seller coach a specific answer and see the
result immediately.

## Posture

Each brief is answered with a strategy posture:

| Posture      | Behavior                                                                |
| ------------ | ----------------------------------------------------------------------- |
| `hold_value` | Protect price. Anchor on defaults and give little discount.             |
| `win_share`  | Compete for the business, within the discount cap and above floors.     |
| `pass`       | Decline, with the reason and the condition under which it would answer. |

## Start by interviewing it

In the app, Semi can interview the seller about the business, then offer a practice brief ("See my
agent handle a brief"). The practice answer appears in the demand inbox as an evaluation, never as
live demand, so you can grade it before any buyer sees your agent.