AgentCore
Building blocks to run agents in production, from any framework and any model. Each piece is adopted on its own and billed on what it uses.
The agent runs in Runtime. Tools go through one Gateway.
MCP is the protocol agents use to call tools. A microVM is a small virtual machine isolated from every other session.
AgentCore runs the agent you already wrote.
Strands, LangGraph, CrewAI
The same Runtime, Gateway and Identity behind each framework. The agent code stays yours.
Bedrock or elsewhere
Models are called from your agent code, inside or outside Bedrock. Tokens are billed by the model provider.
One piece at a time
Already running agents on your own containers? Start with Gateway, Identity or Observability alone.
What decides whether it holds in production.
How sessions run
Runtime bills the CPU and memory a session actually uses. CPU stops billing while the agent waits on a model or a tool.
Which tools, for whom
Put tools behind one Gateway, decide per tool who may call it, and let the agent act with the user's credentials instead of a shared key.
How much to evaluate
Online evaluation scores a share of live sessions. A small sample usually shows drift and keeps evaluation tokens in check.
The evaluation bars are illustrative shapes, not prices.
An AgentCore bill, in one line.
Agentic Platform Foundation
Platform engineering for agents: one foundation, so each team ships its use case faster.
Read the approach →The one-page AgentCore field guide.
Best practices, cost traps and performance checks, in the order you should run them. Printable, no fluff.