21 million exceptions a day: getting Bedrock and AgentCore right

BLP Digital's agentic platform resolves more than 21 million exceptions a day across 550+ enterprise customers in 40+ countries. copebit builds and runs the AWS foundation underneath it. This talk walks both halves.

First, the production system. Rather than retrieving over documents, BLP mirrors each customer's ERP — master data, business rules, process logic — into a Digital Twin, and lets 300+ pre-trained agents act against that typed model through 30+ standard integrations. The model is chosen per processing step on Amazon Bedrock: Anthropic's Claude Sonnet and Claude Opus models, and Gemma alongside them, with all inference in European AWS regions.

Then the part that transfers to your own build. From the AWS foundations we build for Swiss enterprises and the agent platforms we run on Amazon Bedrock and Amazon Bedrock AgentCore, we go through the four decisions that determine whether an agent platform survives a security review: how user identity propagates through to the target system, how large a tool catalog you can afford in context and in cost, how endpoints stay stable, and how registering a new tool becomes a reviewed change.

Plus the failure modes behind each.