Agentic Platform Foundation

Agents change how a company operates, with its customers and inside. A bank thinks about conversational banking: an agent that knows the customer and can act in the core banking system. A manufacturer thinks about an after-sales agent that reads machine telemetry, diagnoses the fault and orders the spare part. Inside, it can be an assistant for an insurer's sales force, or a tool that calculates claim settlements. Each of these use cases can be built on its own. Built that way, the third one is usually where it slows down:

• -Each use case rebuilds its own connections to legacy systems, its credentials and its authentication.
• -Nobody can say which agent can call which system, on behalf of which user.
• -Quality is checked in a demo before launch, then nobody measures it in production.
• -Agent costs arrive on one AWS bill, with no split per use case.

• It is platform engineering, applied to agents. A platform team owns the foundation, and product teams own their use case and start from what already runs.
• We start from the use cases you already run. We keep what the next one will need instead of designing a platform in advance.
• Build vs buy is decided per layer, in writing and with costs: a managed AgentCore service where it covers the need, open source where you need control, a SaaS agent when the use case is generic.
• The agent acts with the user's permissions. Identity flows from your identity provider down to each tool call, not through a service account that sees everything.
• The foundation does not depend on the agent framework. Strands, LangGraph or CrewAI, each team keeps its choice.

1. 01

   Legacy systems as tools

   Core banking, ERP, databases, internal APIs. For those who don't know it, MCP is the protocol agents use to call tools. A single AgentCore Gateway turns your Lambda functions, OpenAPI specs and existing MCP servers into tools, so each system is connected once, then reused by every agent.

2. 02

   Project blueprints

   Each new use case starts from a template architecture: AgentCore Runtime, RAG on your documents, memory per user, and the connectors to the systems it needs upstream and downstream. Teams spend their time on the business logic.

3. 03

   Shared skills and knowledge

   What every agent should know is written once and shared: the approved tone of voice for customer-facing agents, product rules, compliance wording.

4. 04

   Central guardrails

   A dedicated team owns the rules that apply to every agent: Bedrock Guardrails for content and personal data, AgentCore Policy for which group can call which tool, AgentCore Identity for the user's permissions down to each call.

5. 05

   Evaluation, observability & cost

   Every model and tool call is traced in CloudWatch. AgentCore Evaluations scores a sample of live sessions, with alarms when a use case drifts. Cost is split per use case and per team.

• -A 2-hour discovery on the agents you already run and the next use cases on the list.
• -Within a week, a build vs buy memo per layer, costed.
• -Your existing use case moved onto the foundation, then the next one built on it.