AWS Agentic AI Competency
An in-flight engagement within our delivery network — a 50+ agent Amazon Bedrock AgentCore deployment spanning a dozen-plus departments and several channels for a large multi-department enterprise.
The customer is a large, multi-department enterprise building what it calls an AI-native operating layer for customer-facing and internal operations. Details below are drawn from the current architecture and deployment plan, with the customer's identity and sector withheld by agreement: this is a proposal-and-build-in-progress rather than a completed, measured case study.
A high-volume enterprise handles a large number of customer and partner enquiries a year across a dozen-plus departments — marketing, sales, compliance, operations, finance, and more. Manual triage across that many departments creates bottlenecks at any real scale.
A significant after-hours response gap represented measurable revenue leakage, particularly for enquiries arriving outside standard business hours. Eliminating that gap — not just answering faster during business hours — was the stated objective.
Multi-agent orchestration built to serve a dozen-plus departments across several channels within a single AWS region, composed from Amazon Bedrock AgentCore's core services rather than hand-rolled infrastructure.
Web portal, messaging and voice integrations, and an internal design studio for building new agent workflows without a full redeploy.
A master orchestrator and supervisor agents handle routing and classification — deciding which of the specialist agents downstream should handle a given request.
Session-isolated serverless runtime (each agent session runs in its own isolated microVM) paired with AgentCore Identity for OAuth-based, agent-scoped access to downstream tools and APIs — no shared long-lived credentials.
AgentCore Memory holds conversation and task state across turns and sessions; AgentCore Gateway exposes existing CRM/ERP systems and internal APIs to agents as callable tools (including MCP-compatible tool servers) without custom integration glue per agent.
Built-in tracing and metrics (via CloudWatch) across the full multi-agent chain for debugging and audit, backed by a vector-enabled relational store for retrieval and Multi-AZ compute, KMS encryption, and DDoS/WAF protection at the infrastructure layer.
Below a certain agent count, hand-rolled session management works fine on Bedrock Agents directly. Past that point — dozens of orchestration and specialist agents running concurrent, multi-turn sessions across several channels — AgentCore Runtime's per-session microVM isolation, AgentCore Memory's persistent state, and AgentCore Identity's scoped OAuth access stop being nice-to-haves and become the thing that keeps the system operable.
Every stateful service — the database layer, cache, and compute — is deployed Multi-AZ with automatic failover. Shield Advanced and WAF sit in front of the public-facing channels; that security layer alone is treated as non-discretionary for this class of workload.
Figures below describe the provisioned architecture's scale — this is a build-in-progress, so they are scope parameters, not delivered outcomes.
A high-six-figure annualised AWS infrastructure spend across dozens of provisioned services — AI inference (Bedrock/Anthropic models) and compute together account for the majority of that spend, with the remainder split across data, security, networking, and observability.
A wave-based rollout over several months, phasing in departments and channels progressively rather than a single big-bang cutover.
Whether it's a handful of agents or dozens, we'll tell you plainly whether AgentCore or Bedrock Agents directly is the right fit for your scale.
Start a Conversation