2.3.3. Building Agents on AWS: AgentCore, Strands Agents, and Kiro
First Principle: AWS offers agentic AI at three levels: tools for developers building agents (Strands Agents, Kiro), a platform for running agents in production (Amazon Bedrock AgentCore), and finished agentic applications for business users and IT teams (Amazon Quick, AWS Transform).
Amazon Bedrock AgentCore: the platform for running agents
AgentCore is an agentic platform for building, deploying, and operating agents securely at scale with any framework and any foundation model. It works with open-source frameworks such as Strands Agents, LangGraph, CrewAI, and LlamaIndex, and with models in or outside Amazon Bedrock. Its services are modular: use them together or one at a time, with no infrastructure to manage.
| AgentCore service | What it provides |
|---|---|
| Runtime | Serverless hosting for agents and tools; each user session runs in its own isolated microVM; supports long-running work, MCP, and A2A |
| Memory | Short-term session memory and long-term memory that persists across sessions |
| Gateway | Turns existing APIs, Lambda functions, and services into MCP-compatible tools behind one secure endpoint |
| Identity | Agent identities and credentials, including acting on behalf of users (see 5.1.1) |
| Policy | Deterministic rules on which tool calls an agent may make, enforced at the Gateway (see 5.1.1) |
| Code Interpreter | An isolated sandbox where the agent can run code it writes |
| Browser | A managed cloud browser so agents can navigate websites and fill in forms |
| Observability | Step-by-step traces of the agent's reasoning, tool calls, and model interactions |
| Evaluations | Automated assessment of how well agents and tools perform their tasks |
AgentCore also includes newer services, such as a managed Harness (define an agent with a model, a system prompt, and tools in one API call), Payments, Optimization, and Registry. The table covers the services you are most likely to meet at practitioner level.
Strands Agents: the SDK for writing agents
Strands Agents is an open-source SDK, first released by AWS, for building agents in a few lines of Python or TypeScript. It takes a model-driven approach: you supply a model, a system prompt, and a set of tools, and the model plans and decides which tools to call. It supports MCP natively, works with many model providers (including Amazon Bedrock), and includes multi-agent patterns such as swarm, graph, and workflow. A typical path is to write the agent with Strands and deploy it on AgentCore Runtime.
Kiro: the agentic development environment
Kiro is AWS's AI-powered development environment, available as an IDE, a CLI, on the web, and as a mobile app for reviewing work. Its signature is spec-driven development: before code is written, a feature is planned as a spec with three files, requirements.md (user stories and acceptance criteria), design.md (architecture and implementation considerations), and tasks.md (discrete, trackable implementation tasks). Agent hooks run automatically on events such as the agent saving a file, calling a tool, or finishing a task, and steering files give the agent persistent knowledge of project standards. Kiro connects to external tools through MCP.
Amazon Quick: agentic AI for business users
Amazon Quick is a fully managed AI service for work, used through natural-language chat. It evolved from Amazon QuickSight and includes Quick Sight (dashboards and BI), Quick Research (in-depth research delivered as a cited report), Quick Flows and Quick Automate (task and business process automation), Quick Index (grounds answers in the organization's documents), and apps built from a description. It connects to other systems through connectors and open standards, including MCP. No ML expertise is required.
AWS Transform: agents for migration and modernization
AWS Transform uses AI agents to transform infrastructure, applications, and code: migrating VMware and other servers to Amazon EC2, modernizing mainframe workloads, and modernizing .NET applications to cross-platform .NET. It keeps people in the loop: critical actions such as merging to main or deploying to production need approval from a user with the Administrator or Approver role.
Configuration-based agents: AgentCore's managed harness gives a configuration-based starting point: you declare the model, tools, and instructions, and AgentCore handles compute, memory, identity, and observability. It is AWS's recommended path for teams moving off the older Amazon Bedrock Agents (see the status note in 3.1.4).
Scenario: A company has three requests: developers want to build a claims-processing agent and run it in production; the finance team wants an assistant that researches and builds dashboards without coding; and IT wants help moving 400 VMware servers to AWS.
Reflection Question: Which AWS offering fits each request, and why would giving the finance team Strands Agents be the wrong choice even though it can build agents?
💡 Tip: Sort by who is using it. Writing agent code → Strands Agents (and Kiro to write the code). Running agents in production → AgentCore. Business users, no code → Amazon Quick. Migrating or modernizing existing workloads → AWS Transform.
