AI Agent News/stories/amazon-bedrock-agentcore-harness-gets-strands-defaults-built-in-tools-and.md

Infrastructure

Amazon Bedrock AgentCore harness gets Strands defaults, built-in tools and lifecycle hooks

AWS has upgraded Amazon Bedrock AgentCore harness to use the open-source Strands harness, adding managed defaults, new built-in tools and lifecycle hooks. Agents on the managed defaults get a recommended model, a task-oriented system prompt, automatic prompt caching and proactive context management without new resources to adopt.

Source: aws.amazon.com/blogs/machine-learning/amazon-bedrock-agentcore-harness-adds-strands-harness-defaults-built-in-tools-and-lifecycle-hooks

AWS is upgrading AgentCore harness to use Strands harness, its open-source agent harness released earlier this year under an Apache 2.0 license. Agents that rely on the managed defaults get a recommended model that changes as new models ship, a task-oriented system prompt, and automatic prompt caching, which initially applies to Anthropic models on Amazon Bedrock. Proactive context management summarizes older conversation turns once the context window reaches a configurable threshold, and AWS says this keeps token cost about 34% lower than other leading harnesses.

The upgrade adds built-in tools for planning with todo_write, web retrieval with web_fetch, and separate read, write and edit filesystem tools. It also offloads large tool results from the model context and includes a subagent tool that hands a self-contained task to a fresh worker. Setting a model ID, systemPrompt, truncation or an explicit allowedTools list keeps a fixed configuration. AWS began rolling out the change on November 11 and says existing configurations are not affected.

Lifecycle hooks run deterministic code at four points in the agent loop: before_invocation, before_tool_call, after_tool_call and after_invocation. Only Lambda hooks can allow or deny a step, while Amazon SNS and Amazon EventBridge hooks only send notifications. If a Lambda hook times out or errors, its failureMode applies, which defaults to deny.

Why it matters if you build agents

Agents on the managed defaults will pick up new models and tools over time, so teams with a fixed security or capability boundary should set a model and list only the allowedTools they need. A before_tool_call Lambda hook can gate consequential actions such as refunds before they run. Subagent tool calls pass through the same hooks, so delegation does not bypass those checks.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe