Home / Blog / Amazon Bedrock AgentCore harness adds Strands harness…
Tech News

Amazon Bedrock AgentCore harness adds Strands harness defaults

Amazon Bedrock upgraded AgentCore harness to open-source Strands harness defaults with built-in tools, context pruning, and deterministic lifecycle hooks.

By Dillip Chowdary • Nov 19, 2026 • Source: AWS Machine Learning Blog

Amazon Bedrock AgentCore harness adds Strands harness defaults
Image: AWS Machine Learning Blog

At a glance

Product
Amazon Bedrock AgentCore harness
Open-source foundation
Strands harness
License
Apache 2.0
Pricing advantage
~34% lower token cost
Default failure mode
Deny
Auto-approval limit
$500 for refund example

Amazon Web Services upgraded its Amazon Bedrock AgentCore harness to run on the open-source Strands harness, introducing recommended model defaults, built-in productivity tools, proactive context pruning, and deterministic lifecycle hooks. According to AWS Machine Learning Blog's report, the release brings production-grade agent configurations directly into existing harnesses under an Apache 2.0 license. The runtime upgrade requires no migration, new resource types, or manual infrastructure provisioning for existing users.

This report details the architectural transition to Strands, the expanded tool surface, context caching mechanics, security configurations, and programmatic governance controls for enterprise engineering teams. Developers running autonomous AI workloads on AWS can now deploy isolated agent loops while establishing strict guardrails around consequential operations. The update provides software architects and platform engineers with verifiable execution safety, lower token costs, and automated session persistence across multi-turn interactions.

01Managed microVM execution and runtime architecture

The foundation of the upgraded service relies on Strands harness, the open-source agent framework that AWS released earlier this year under an Apache 2.0 license. Instead of requiring developers to package agent loops or administer custom execution servers, Amazon Bedrock operates the framework as a fully managed cloud service. Every active session runs inside a secure, kernel-isolated microVM within the AgentCore runtime to ensure robust environment isolation, automated scaling, and continuous high availability.

The managed runtime connects the agent loop directly into foundational AWS services necessary for sustained enterprise operations. These automated integrations provide out-of-the-box session persistence, streaming response generation, granular usage metadata tracking, comprehensive observability metrics, and AgentCore memory. The runtime also hooks directly into AgentCore Gateway, Identity access controls, Browser automation, the Code Interpreter sandbox, and customer-provided container environments without manual wiring.

02Built-in tools and prompt caching mechanics

When developers omit explicit model selections, the harness dynamically supplies a recommended foundation model that automatically updates as newer options debut. The managed system prompt directs the agent to gather relevant context, adapt gracefully to denied tool actions, and verify results before signaling completion. For Anthropic models interacting through Converse on Amazon Bedrock, automatic prompt caching stores system prompts, tool definitions, and conversation histories to reduce input processing latency and billing overhead.

Context growth in long conversations is addressed through automated summarization and offloading mechanisms designed to preserve memory efficiency. Once conversation history crosses a configurable threshold, the harness compresses older dialogue turns while keeping recent context intact, achieving approximately 34% lower token cost than other leading harnesses. Large tool outputs are also moved out of immediate model context into on-demand previews, while native tools like todo_write, subagent, web_fetch, and granular filesystem commands handle execution.

03Explicit configuration and tool security boundaries

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Although the managed harness delivers optimized defaults that update automatically, development teams can enforce strict deterministic configurations whenever standardized behavior is required. Setting an explicit model identifier locks execution to a chosen model version, while supplying a custom system prompt completely replaces the default task-oriented instructions. Teams can similarly override context summarization thresholds by configuring custom truncation rules that govern retention policies during long-running operational workflows.

Note: Configuring the allowedTools parameter allows teams to restrict an agent capability surface to exact tools like read and web_fetch while disallowing file modifications or shell access. Automatic prompt caching is initially available for Anthropic models on Amazon Bedrock using Converse, while existing harness deployments inherit the broader Strands capabilities automatically without requiring new AWS resource types.

04Four lifecycle events and synchronous enforcement

Deterministic control over autonomous agent execution is established through lifecycle hooks placed at critical operational checkpoints. Engineers can attach synchronous AWS Lambda functions to evaluate execution state, enforce compliance boundaries, and return explicit allow or deny decisions. Asynchronous targets like Amazon Simple Notification Service topics and Amazon EventBridge event buses can also subscribe to these checkpoints to emit audit logs or trigger decoupled downstream workflows without blocking agent progress.

| Event | When it runs | Effect of a Lambda denial | | before_invocation | Before the agent processes the request | Stops the invocation with hook_stopped | | before_tool_call | After the model proposes a tool call and before execution | Skips the tool call. The agent loop continues | | after_tool_call | After a tool result exists | Retains the result and stops before further inference | | after_invocation | After the invocation completes | Reports the decision. Output already streamed cannot be retracted |

Only Lambda functions possess the programmatic authority to alter agent trajectory during an active run. When invoked synchronously, the Lambda function evaluates the incoming hook payload and returns a structured decision containing an allow or deny directive alongside an optional explanatory reason. Notifications routed to Amazon SNS or EventBridge remain non-blocking, ensuring monitoring pipelines do not introduce artificial latency into real-time user interactions.

05Subagent policy delegation and failure modes

Real-world policy enforcement can be illustrated by a customer-support agent handling financial transactions. In this setup, a before_tool_call hook inspects an issue_refund tool call against a strict $500 auto-approval limit. If the proposed refund exceeds that limit, the Lambda function returns a denial with an explicit reason, prompting the harness to skip execution while allowing the agent to inform the customer that the request was escalated.

Policy controls extend seamlessly across distributed workflows when an agent delegates work through the subagent tool, with child workers executing under identical beforeToolCall and afterToolCall validation hooks. Every programmatic decision surfaces directly in the InvokeHarness streaming response as a distinct hookEvent object for real-time monitoring. If a hook Lambda function times out or encounters an unhandled runtime error, the harness defaults to a strict failureMode of deny, preventing uninspected actions from executing.

Frequently asked questions

What is the Strands harness in Amazon Bedrock AgentCore?

Strands harness is an open-source agent framework released by AWS under an Apache 2.0 license that provides managed defaults, context pruning, and built-in tools.

How do lifecycle hooks govern autonomous agent actions?

Synchronous AWS Lambda hooks evaluate requests at four checkpoints to allow or deny actions like issuing refunds, while Amazon SNS and EventBridge receive non-blocking audit notifications.

How does proactive context management reduce token expenses?

The harness compresses older conversation turns above a threshold and offloads large tool results into on-demand previews, cutting token costs by approximately 34%.

Do security hooks apply when an agent delegates tasks to subagents?

Yes, subagents share invocation limits and automatically inherit the parent agent's beforeToolCall and afterToolCall lifecycle hooks.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →