AI‑Native commerce run by agents
Period — Present
deepeet started as a consignment business selling fruit from wholesale suppliers on sales channels. The first piece was order automation, which turns sales‑channel orders into purchase orders with wholesale suppliers. Routing, pricing, settlement, sourcing, listings, customer support and advertising followed, each built in‑house on top of it. AI agents decide from evidence, and the commerce system executes only after checking permissions and safety conditions. Each result is then checked against the official status at the sales channel or supplier. Important code changes are verified by a different model from the one that wrote them.
Reached 700+ orders and over ₩10M in revenue within three months through automation.
Completion reports versus actual state
- Problem
- An agent's completion report can differ from the actual system state.
- Solution
- The lead reviews the report, and the system executes only procedures that pass the rules kept in vat. The actual state is checked again afterwards. All 19 underlying repositories follow the same rules.
Architecture
Single point of contact for the user · task assignment · result checks · final decisions
Knowledge and memory
- Business records
Review before changes
- Lead agent records
Principles · workflows · operations journal
- Conversation memory
Summarized conversations and user information
- Search index
Search material derived from records · not used as evidence
Execution and control
- Commerce systems
Order automation · routing · pricing and settlement · sourcing · listings · operations console · advertising
- Task board
Long‑running work, handovers and retries
- Release process
Changes across 19 repositories · validation of what changed · checks after deployment
- Permissions
L0 lookups, classification and drafts · L1 running defined procedures
- Unattended operation
Messenger connection monitoring · automatic start after reboot · encrypted backups
Development
- InvestigationL0
Root causes and technical evidence
- ImplementationL0
Work within assigned tickets
- ReviewL0
Quality, regression and security checks
Business operations
- ProductL1
Planning, detail pages, listings and verification
- Customer supportL0
Classifying text inquiries and drafting replies
- FinanceL0
Source lookups, reconciliation and reports
- MarketingL0
Advertising and marketing decisions
- Supplier monitoringL0
Checking and organizing supplier notices
Agent workResult reportLead reviewSystem action after permission checksActual state check