Agent Platform Stack
Architects evaluating how the AgentCore platform layers compose to support different agent workloads — the horizontal layers, vertical pillars, and team consumption models.
The agent platform stack is a layered architecture that provides shared capabilities to domain teams. Each layer handles a specific concern, while cross-cutting pillars ensure security, governance, and operational excellence across all layers.
Core Platform Layers
The horizontal layers form the runtime backbone of the platform, spanning everything from request ingestion to operational feedback loops.
AI Gateway
Unified entry point for all agent interactions. Handles routing, auth, rate limiting, and protocol translation.
Registry & Discovery
Central catalog of all agents, their capabilities, versions, and ownership. Enables discovery and prevents duplication.
Agent Runtime
Sandboxed execution environment for agents. Isolation, scaling, long-running task support, and resource management. Runtime Instances extend this to AWS-managed EC2 for long-running, GPU-backed, or persistent-disk workloads.
Memory & Context
Short-term and long-term memory, vector stores, knowledge bases, and session context management.
Observability
Distributed tracing, metrics, logs, and audit trails across all agent executions and tool invocations.
AgentOps
CI/CD, evaluations, A/B testing, rollback, and continuous quality assurance for the agent fleet.
Security Pillar
Security is not a layer - it's a vertical pillar that intersects every platform component.
| Capability | Description |
|---|---|
| Agent Identity | Each agent has a unique, verifiable identity for AuthN/AuthZ |
| Access Control Policies | Fine-grained, declarative policies governing what agents can access |
| Tool Permissions | Scoped permissions for each tool an agent can invoke |
| Guardrails | Input/output filtering, PII redaction, topic denial, hallucination detection |
| Data Protection | Encryption at rest and in transit, data classification, retention policies |
| Shadow AI Detection | Identify and flag unregistered agents operating outside the platform |
Governance Pillar
Governance ensures agents operate within organizational policies throughout their lifecycle.
| Capability | Description |
|---|---|
| Lifecycle Management | Creation → approval → deployment → deprecation → retirement |
| Policy-as-Code | Declarative policies versioned and enforced automatically |
| Compliance | Automated checks against regulatory frameworks (EU AI Act, NIST, ISO 42001) |
| Audit Logs | Immutable record of all agent actions, policy decisions, and configuration changes |
| Shadow AI Prevention | Governance workflows that make the "right path" the "easy path" |
Platform Ops
Platform operations ensure the platform itself runs reliably and scales with organizational growth.
| Capability | Description |
|---|---|
| Platform CI/CD | Infrastructure-as-code pipelines for platform components |
| Tenant Onboarding | Self-service workflows for new teams to start building agents |
| Quotas & Limits | Per-team resource quotas, token budgets, and rate limits |
| Multi-Account Management | AWS Organizations, SCPs, and account vending for isolation |
| IaC Constructs | Reusable CDK/CloudFormation constructs for common patterns |
| Cost Tracking & Attribution | Per-agent, per-team, per-use-case cost visibility |
How Domain Teams Consume the Platform
The platform serves teams with different needs and maturity levels. Three common consumption models:
Team 1 - Consume
Uses pre-built agent templates and platform services. Configures agents through the registry, deploys via self-service pipelines. Minimal custom code.
Low EffortTeam 2 - Build on Platform
Brings their own agent code and orchestration logic. Uses platform services for runtime, memory, observability, and governance. Custom agents, shared infrastructure.
Medium EffortTeam 3 - Build Independently
Builds with their own framework and infrastructure but registers agents in the platform for discovery, governance, and observability. Federation model.
Higher EffortThe best platforms make Team 1 the default path and Team 3 the exception. The easier it is to consume the platform, the less Shadow AI you'll encounter.