AWS details AgentCore multi-account architecture splitting inference and tool gateways
AWS published a reference blueprint for AgentCore that separates inference from tool execution across account boundaries and isolates each workstream in its own cell. Agent deployments are gated by a CI/CD pipeline that measures accuracy, hallucination rate, latency and cost before promotion.
- Inference gateway runs in a shared Bedrock account; tool gateway lives in a per-workstream cell
- Control plane routes requests and SCPs block direct cell-to-cell calls
- Accuracy threshold is 95% correct tool calls; hallucination rate capped at 2%
- p99 latency of 5 s and $0.10 per request trigger warnings, not blocks
Read next
AI