Multi-Agent AI Team for Sales, Marketing and Ops with Human Approvals
Our own operations (RetailBox)
Retailbox: our own operations, 2026
Our own sales, marketing and ops run largely on a self-hosted agent platform, Paperclip: 34 agents hold roles such as head of sales, SDR, CMO, SEO and DevOps, and people approve what goes outside.
What was the problem?
Agents that act for a business need clear roles, someone to escalate to, and limits on what they can do outside the company.
How did Retailbox solve it?
34 agents hold roles like head of sales, SDR, CMO, SEO, DevOps and tender manager, each with written instructions and a manager to escalate to. Outward actions go mostly through Windmill robots with limits: LinkedIn posts publish after a human approval card, website changes merge through an approval gate with a live-link check, and marketing email is capped at 10 a day. Spending caps, messages to existing clients and new access go to a person.
Client project work is done by people and by separate Claude Code sessions, not by this team. Some agents also fill in web forms directly, so not every outward action goes through a robot. The task count below includes recurring duty and alert tasks.
Facing the same problem? These are the services that cover this kind of work: AI agent development services.
What does it look like?
(opens the full-size image in a new tab)What were the results?
- ✓34 agents on the platform; 32 of them closed tasks between October 21 and November 20, 2026
- ✓6,137 tasks closed by agents from October 21 to November 20, 2026, duty and alert tasks included
- ✓LinkedIn posts publish only after a human approval card
- ✓Website changes merge through an approval gate; marketing email is capped at 10 a day
- ✓Spending caps, messages to existing clients and new access go to a person