01HARNESSBUILD
The easiest way to run an agent: describe it in config (model, instructions, tools) and AgentCore runs the whole loop for you, no orchestration code.
AgentCore is a set of AWS services for building, running, and operating AI agents. An agent is a program that uses an AI model to take actions, not just chat. You can use any agent framework (Strands, LangGraph, CrewAI, LlamaIndex, Google ADK) and any model. The services work together, or one at a time.
New to these words? Chapter 02, "Key Terms", defines them all: agent, tool, model, session, microVM, MCP, A2A.You create the agent: pick a model, write its instructions, and give it tools (functions and APIs it can call). Use the Strands framework, or let Harness manage it.
Runtime runs the agent for you, with no servers to manage. Each session runs in a private, isolated space (a microVM).
Memory stores the conversation and the facts the agent learns, so it does not forget between turns or sessions.
The agent does real work by calling tools: your own APIs through Gateway, plus a built-in web Browser and a Code Interpreter.
Using A2A (Agent-to-Agent), your agent can hand part of the job to other agents.
The easiest way to run an agent: describe it in config (model, instructions, tools) and AgentCore runs the whole loop for you, no orchestration code.
Serverless hosting for your agent. It scales by itself and runs each session in its own isolated microVM, so you focus on the agent, not servers.
Gives the agent a memory. Short-term holds the current chat; long-term keeps useful facts and preferences for later. Without it, the agent starts blank every time.
Turns your existing APIs and AWS Lambda functions into tools the agent can call, behind one secure address, so you skip writing custom code for each tool.
An AWS-managed, locked-down sandbox where the agent runs real code (Python, JavaScript, TypeScript) to compute or analyze data, safely away from your systems.
An AWS-managed cloud web browser the agent can drive to open sites, fill forms, and read pages, with a recording you can watch back (session replay).
Controls who may use the agent and what the agent may access. It works with your existing login providers (Cognito, Okta, Microsoft Entra ID) and safely holds access tokens.
Rules you write (plain language or AWS Cedar) that check and allow or block every tool the agent tries to use, keeping it inside your limits.
A shared catalog where teams publish and find approved agents, tools, and MCP servers, with search, so people reuse vetted parts instead of rebuilding them.
Lets you see what the agent did, step by step (traces and spans), in Amazon CloudWatch, to debug problems and watch quality.
Automatically scores how well the agent does its job, using AI judges and your own checks, so you catch problems before and after launch.
Continuously improves the agent with AI-generated recommendations, versioned configuration bundles, and A/B testing, so quality goes up instead of drifting down.
Lets the agent pay tiny amounts of real money to use paid APIs or content, with spending limits, using the open x402 standard. This is real money, not LLM tokens.
How an agent finds and calls tools. This is the language Gateway speaks.
How one agent asks another agent to do part of the job, even across different companies. An open standard (Linux Foundation).
A session is one continuous run with a user, named by a session id (reuse it to continue the same conversation). Each session runs in a microVM, a tiny private computer just for that session, which is wiped when the session ends so nothing leaks between users. A session can last up to 8 hours. AgentCore runs on ARM64 (AWS Graviton) chips.