Axonera: An AI Agentic Operating System Built for Retail
A five-layer AI agent PaaS — SDK, invocation kernel, CQRS API, PostgreSQL, and a Next.js portal — for teams building AI agents that need to run reliably in production, not just demo well.
The business problem
Every AI agent framework on the market solves the easy 20% — the demo. The remaining 80% is what enterprises actually need: multi-tenant isolation that survives operational mistakes, invocation modes that don’t hold HTTP connections open for hours, workflows a non-engineer can inspect, and durable audit trails. Off-the-shelf agent orchestration doesn’t cover it. Hand-rolling it for every client engagement is a tax we didn’t want to keep paying — so we built the platform once, properly.
What we delivered
- A five-layer platform: TypeScript SDK for agent authoring, an in-process invocation kernel with four modes (sync, async, streaming, batch), a CQRS-shaped API surface, PostgreSQL persistence with row-level security for multi-tenant isolation, and a Next.js portal for workflow inspection.
- Multi-tenant row-level security baked into the data layer — not sprinkled across application code, where mistakes leak data across tenants.
- Native authentication (no third-party dependency) with SSO wiring for the Azure AD tenants most of our retail clients already run.
- Retail-vertical opinionation: connectors, ontologies, and workflow templates tuned to how retailers actually run agent-augmented operations — merchandising, inventory, customer support automation.
Results
Multi-tenant by design
PostgreSQL RLS enforces tenant isolation at the database, not at the application — the only pattern that’s safe under mistakes.
Four invocation modes
Sync, async, streaming, batch — chosen per workflow so long-running agents don’t hold HTTP connections open.
Portal-first UX
Every workflow is inspectable and editable in the Next.js portal — non-engineers can see what an agent did and why.
Retail-native
Connectors, ontologies, and templates that reflect how retailers actually run — not a generic agent framework retrofitted onto retail.
Fix the data warehouse. Then build AI on top.
Twenty minutes to talk through your Snowflake environment or your AI plans — whichever is holding you up. No pitch, no pressure.