aws-strands-agents-agentcore_skill

Overview

This skill helps teams design, implement, and operate AI agents using the AWS Strands Agents SDK and the Amazon Bedrock AgentCore platform. It captures architecture patterns, deployment decision matrices, observability and evaluation practices, multi-agent orchestration patterns, and MCP server integration rules. The content targets production-ready agent systems with concrete rules, templates, and anti-patterns to avoid costly mistakes.

How this skill works

The skill inspects your agent topology and recommends patterns (single agent, agent-as-tool, graph, swarm) and deployment targets (Lambda, ECS/Fargate, AgentCore Runtime) based on interaction style, state needs, and streaming requirements. It enforces MCP server constraints (streamable-http, specific endpoint, ECS/Fargate or AgentCore deployment), guides tool design and token management, and provides observability and evaluation integration steps for continuous quality monitoring.

When to use it

• Designing or refactoring agents that will run on AWS Bedrock or Strands SDK
• Selecting deployment targets for stateless, interactive, or stateful agent workloads
• Building MCP tool servers that require persistent connections and a tool registry
• Implementing observability, session management, and production evaluations
• Planning multi-agent orchestration or scaling strategies

Best practices

• Start with a single agent and introduce multi-agent patterns only when needed
• Keep tool count low; use semantic search/dynamic loading when >15–50 tools
• Never deploy MCP servers to Lambda; use ECS/Fargate or AgentCore Runtime with streamable-http
• Implement conversation managers (sliding window or summarisation) to control tokens and costs
• Instrument agents with OpenTelemetry and enable AgentCore Observability and Evaluations for production

Example use cases

• Stateless API agent: deploy as Lambda for short, deterministic requests
• Interactive chat agent: run on AgentCore Runtime for streaming and cross-session memory
• Tool/Integration MCP server: run in ECS/Fargate with persistent DB and streamable-http transport
• Deterministic workflow: implement a graph pattern for predictable multi-step processing
• Autonomous collaboration: use swarm pattern for multiple agents coordinating complex tasks