Daniel Wirjo

A SaaS control plane - tenant onboarding, billing, configuration management, agent provisioning is the most repeatable software every software vendor builds from scratch.

Agentic workloads are multi-step, tool-calling, and retrieval-augmented, which means they stress the inference layer differently than a single prompt. In this code talk, we build three patterns live: function calling, multi-step reasoning chains, and retrieval-augmented generation with Knowledge Bases.

Managing memory for hundreds of tenants in SaaS shouldn't require hundreds of IAM policies — yet most products scaling multi-tenant agents on AWS end up there: a namespace per tenant-agent pair, a policy per tenant, deployment times that climb with every customer.

Agents are meant to be autonomous and intelligent. Ensuring that this intelligence is operating as expected, requires all new approach to evaluating and testing our agentic systems.

SaaS is meant to be a data-driven model. The choices you make when architecting a multi-tenant system are driven by specific metrics, goals, and outcomes that are expected to enable growth, innovation, and efficiency.