Reactiv's AI Scheduler cuts configuration time 80 percent on Amazon Bedrock AgentCore

Reactiv's move to Amazon Bedrock AgentCore was driven by four specific gaps in its earlier setup: no multi-agent orchestration for scheduled tasks, no cross-session memory, a custom authentication and JSON-RPC handshake layer for its Config MCP, and roughly 100 OpenAPI spec files maintained across two locations. The company had already built a conversational AI builder for merchants to modify apps in real time, but merchants wanted autonomous updates on a schedule without manual intervention. That gap led Reactiv to package its Strands agent graph as a Docker image and run it on AgentCore, where agents spin up on a schedule and shut down when done.

After launch, Reactiv faced a second problem: its interactive dashboard agent and scheduled agent were separate systems that shared no memory, tools, or infrastructure. Unifying them on the AG-UI protocol on AgentCore produced bidirectional memory sharing, so preferences learned during a dashboard session feed into the next scheduled run, and vice versa. The architecture also relies on per-merchant Firecracker microVMs for multi-tenant isolation, keeping each merchant's execution context and memory separate.

Looking ahead, Reactiv is extending the authoring surface for merchants: making its existing design properties available to the agent as a new MCP server hosted on AgentCore, rebuilding onboarding so a new merchant's brand name and website produce a starting app, and eventually letting merchants request entirely new layout sections through natural language. The stakes for Reactiv hinge on whether the governed design vocabulary and per-merchant memory can keep autonomous updates on brand, and whether the internal time and cost savings hold as the merchant base grows. For other Shopify app builders, the pattern of hosting MCP servers on a managed runtime to share tools across agents may be worth watching.