Topic overview
Strands takes a model-driven approach. Instead of hand-drawing a workflow, you give an agent a model, a system prompt, and tools, and the model plans, calls tools, reads the results, and repeats until it can answer. Because modern models are good at this reasoning loop, a working agent can take only a few lines of code. It works with Amazon Bedrock by default and with other model providers as well. Framework details change quickly, so check the official docs for current APIs.
Core concepts
The agent loop: the model receives the task and the available tools, decides what to do, calls a tool, sees the result, and continues until it produces a final answer.
Tools: ordinary Python functions turned into tools, typically with a decorator that uses the function name, docstring, and type hints as the tool description and schema.
Model providers: Amazon Bedrock is the default, and other providers such as Anthropic, OpenAI, and local models can be used, so the agent code stays the same.
MCP support: an agent can use tools exposed by Model Context Protocol servers, which gives it a large catalogue of ready-made integrations.
Multi-agent patterns: agents can be used as tools by other agents, or arranged as a swarm, a graph, or a workflow when you want more structure than a single loop.
Conversation state and memory: the agent keeps message history across turns, and you can manage how much context is kept.
Observability: built-in tracing with OpenTelemetry makes it possible to see model calls, tool calls, and latency in your usual monitoring tools.
Deployment: agents are normal Python programs, so they run on AWS Lambda, containers, or Amazon Bedrock AgentCore, and equally well elsewhere.
Why it matters
Not every agent needs a hand-built graph. A model-driven SDK shows how far you can get with a good model, clear tools, and a clear prompt, and it is a natural fit if your stack is already on AWS. Knowing it alongside LangChain and LangGraph lets you pick the right level of control for the job.
Interview relevance
Expect comparisons and design questions: model-driven versus workflow-driven agents, when a single loop is enough and when you need explicit control, how you would expose a company API as a tool, how you would add guardrails and observability, and how you would deploy and scale an agent on AWS. A good answer states the tradeoff: less code and more flexibility, but less predictable control flow than an explicit graph.