Strands

Amazon Bedrock AgentCore is a managed service launched by AWS in 2025, designed specifically for running and orchestrating AI agents in production environments.

When review requires additional research, Strands discards the 306 facts already checked and starts over, resulting in a single-run cost as high as $11.81.

Strands sets the maximum number of executed nodes via set_max_node_executions(), and once exceeded the graph state becomes FAILED.

Strands adds a hook to each Agent via add_hook to check cumulative cost before model calls; ADK achieves similar functionality via before_model_callback and after_model_callback, and supports a plugin mechanism so that one registration applies to all Agents.

Strands wraps exceptions into EventLoopException after tool execution; ADK first emits an error event and then throws the exception at the end of the stream.

Real cost data shows that building the Terraform course with Strands cost $18.36, while ADK cost only $3.35 (excluding upfront research and planning).

Strands itself does not provide search tools and requires connecting to an MCP server; ADK has built-in google_search but only for Gemini models.

Strands uses MCPClient to connect directly to the AWS documentation MCP server, while ADK integrates it through McpToolset.

In terms of page reading, Strands' web_fetch supports markdown mode to return the original text, making it easy to verify facts word by word; ADK's load_web_page has problems with redirects and short-line filtering, causing large numbers of code examples to be lost.

Strands passes structured_output_model= when calling, and the result is obtained via result.structured_output, directly as a Pydantic model instance.

In Strands, graph nodes must be Agent or MultiAgentBase subclasses, and ordinary functions will throw a ValueError at runtime.

Using Strands' BedrockModel to connect to Bedrock requires only a small amount of code, whereas SageMaker requires three API calls—create-model, create-endpoint-config, and create-endpoint—and in testing the endpoint took about 9 minutes to deploy to InService.

Compared to frameworks such as LangGraph and AutoGen, Strands operates at a higher level of abstraction with less code, but relies more heavily on the reasoning capabilities of the underlying model.

The original article uses a supply chain decision-making scenario as an example to build a multi-agent system with the Strands framework, with Strands responsible for defining each agent's role, the set of callable tools, and the collaboration process.

Strands' design philosophy is model-driven: developers only need to declare tools and constraints, and the underlying model autonomously determines the invocation order and reasoning path, rather than hardcoding execution steps.