Serverless AI agents, without a rewrite

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The migration centred on three areas: elastic and cost-efficient compute, a zero-disruption cut-over with decision parity, and preserving the existing agent codebase.

Cloud Combinator ran the migration across six phases, consolidated below, keeping the existing ECS deployment fully operational until an intentional, validated decommission at the end.

Acceptance was structured around sign-off of the success criteria, phase-level acceptance at each of the six stages, and a production cut-over executed with minimal disruption, rollback capability and no data loss. The figures below reflect the SoW volume assumptions and projected cost model.

• Serverless compute — AgentCore Runtime's scale-to-zero, scale-on-demand model replaces always-on ECS Fargate, so compute cost tracks demand.
• Zero rewrite — The Mastra framework, 14 agents, 18 tools and 2 orchestration workflows migrate unchanged.
• Validated parity — A week of shadow-mode comparison and a staged canary-to-full traffic shift, with a tested rollback to ECS, de-risk the cut-over.
• Observability built in — CloudWatch dashboards and alerts plus X-Ray tracing give the team visibility into the new runtime.
• Handover — Deployment and rollback runbooks, an operations guide and a workshop covering AgentCore operations, scaling behaviour and incident response were delivered.

With the DLP email security service now running on AgentCore Runtime, the client's compute footprint scales elastically with demand rather than being fixed and always on. The target architecture is designed for full multi-region active/active operation, which can be enabled in a follow-on phase.