Amazon Bedrock AgentCore harness updates add Strands defaults and lifecycle hooks

Amazon Bedrock has upgraded its AgentCore harness to incorporate the Strands agent foundation, introducing automated context management, built-in tools, and deterministic lifecycle hooks.

2 min read

Amazon Bedrock AgentCore harness updates add Strands defaults and lifecycle hooks

Building autonomous AI systems often requires assembling multiple moving parts beyond the core model itself. Developers must configure tools, write system instructions, manage growing context windows during extended sessions, and enforce safety guardrails before an agent executes consequential actions like financial transactions. To address these operational hurdles, Amazon Bedrock has upgraded its AgentCore harness to utilize Strands harness, an open-source framework released under an Apache 2.0 license.

Key facts

Python Version
v0.4.0
TypeScript Version
v0.4.0
Release Date
Nov 12, 2026

Strands harness has 34% lower estimated costs at the median across matched Claude and GPT runs

Product demonstration

Benchmark comparison illustrating the estimated cost savings of the Strands harness foundation.

Watch full demonstration ↗

Managed defaults and capabilities

By adopting the Strands foundation, AgentCore harness now provides a robust set of managed defaults designed to reduce infrastructure overhead. When developers rely on these defaults, the system automatically assigns a recommended model, task-oriented instructions, and proactive context management. Automatic prompt caching is also enabled initially for Anthropic models on Amazon Bedrock, which helps lower processing latency and cost for repeated calls.

To help agents handle long tasks, the updated harness introduces several built-in tools. These include a planning utility to maintain task lists, a web retrieval tool for external context, granular filesystem access through separate read, write, and edit tools, context offloading for large results, and a delegation tool for subtasks. According to the update, these optimizations allow the managed harness to run at approximately thirty-four percent lower token cost compared to other leading options.

Governing workflows with lifecycle hooks

In addition to expanded tooling, the release introduces lifecycle hooks that allow developers to execute deterministic code at specific checkpoints in the agent loop. These hooks support four distinct lifecycle events, enabling teams to transform non-deterministic outputs into predictable workflows.

For example, developers can attach an AWS Lambda function to inspect tool calls before execution. If an agent attempts an action that exceeds defined limits, such as a refund request above a specific dollar amount, the Lambda function can evaluate the context and return a decision to deny the call. Other supported targets, such as Amazon Simple Notification Service and Amazon EventBridge, provide non-blocking notifications to record activity or trigger asynchronous workflows without interrupting the primary loop.

Streamlining multi-agent coordination

Creating and managing these environments is handled directly through the updated AgentCore CLI, which allows developers to provision harnesses with a single command. The architecture executes each session within a secure, kernel-isolated microVM in the AgentCore runtime. This isolation handles compute, scaling, and high availability automatically, allowing engineering teams to focus on agent instructions, permissions, and tool surfaces rather than underlying infrastructure management.

Source: Original announcement

Worth a read?

Report

Comments