Amazon Bedrock AgentCore harness adopts the Strands harness: new built-in tools and 4 lifecycle hooks, rolling out from November 11
AWS says Amazon Bedrock AgentCore harness now uses the open-source Strands harness defaults, with new built-in tools and lifecycle hooks. Existing harnesses get them automatically; explicit settings stay unchanged.
News date: About 7 min read
Contents
AWS announced on its official blog that Amazon Bedrock AgentCore harness has been upgraded to use its open-source Strands harness (Apache 2.0 license) and now supports lifecycle hooks. Here, a harness can be understood as the execution framework wrapped around an AI model, which supplies the agent with tools, instructions and context management; hooks are checks or notifications you write yourself that fire automatically at specific points in the agent's run. According to AWS, the experience is rolling out to new and existing harnesses from November 11, with no new resource types or extra integrations required. For teams building AI agents, the managed service can now provide defaults for long-task context management, tool permissions and guardrails on high-risk actions such as refunds.
What changed in this upgrade?
AWS says agents using managed defaults get the currently recommended model (if you omit the model setting, the default model evolves as new models are released), a task-oriented system prompt, automatic prompt caching (reusing already-processed prompt content to reduce latency and cost; initially for Anthropic models on Amazon Bedrock), and proactive context management. AWS states that, for supported models using Converse on Amazon Bedrock, the harness caches the system prompt, tool definitions and conversation history, so repeated calls can use lower-priced cache-read tokens (a token is the unit models use to measure and bill text).
On context management, the context window is the amount of content a model can process at once. AWS explains that by default the harness summarizes older conversation turns when the context window reaches a configurable threshold, keeping recent content; AWS says this makes its token cost about 34% lower than other leading harnesses (this figure is AWS's own claim). Large tool results can be offloaded so the agent receives only a preview and retrieves the full content when needed.
| Built-in tool | Purpose as described by AWS |
|---|---|
| todo_write | Maintains a structured task plan that later calls in the same session can restore, including after summarization |
| subagent | Hands independent tasks to a fresh worker (subagent) and returns results to the parent agent; shares call limits and tool-call hooks |
| web_fetch | Retrieves the content of a specified web page to provide up-to-date external information |
| read, write, edit | Separate filesystem tools, so you can grant only the fine-grained permissions needed |
How do lifecycle hooks guard high-risk actions?
AWS says hooks let you run your own deterministic code (the same input always gives the same result, unlike AI responses, which can vary) at specific points in the agent loop. Lambda functions can allow or deny a request or tool call; Amazon SNS topics or Amazon EventBridge event buses can receive notifications. Only Lambda hooks can change the agent loop, and they run synchronously; SNS and EventBridge are notification-only and do not block.
| 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, before it runs | Skips that tool call; the agent loop continues |
| after_tool_call | After the tool result is produced | Keeps the result and stops before further inference (the model's next processing step) |
| after_invocation | After the invocation completes | Reports the decision; output already streamed cannot be recalled |
AWS's example is a customer-service agent: a before_tool_call hook checks the amount before the refund tool runs and denies it if it exceeds the $500 automatic approval limit, and the agent can tell the customer it has been escalated. AWS adds that hooks also apply to every tool call made by a subagent, so delegation cannot bypass policy. Hooks are configured through CreateHarness or UpdateHarness; if a Lambda times out or errors, failureMode applies, and it defaults to deny.
Rollout and compatibility: what do I need to do?
AWS says the Strands harness experience began rolling out to new and existing harnesses on November 11, with no customer action needed. Settings you provide explicitly stay unchanged; harnesses that omit the model, system prompt or truncation settings, or that use allowedTools: ["*"], get the upgraded managed defaults. AWS says customer-provided MCP servers, AgentCore Gateway, Browser, Code Interpreter, inline functions, skills, memory and bring-your-own containers continue to work through existing interfaces.
| Desired effect | Setting as described by AWS |
|---|---|
| Keep a specific model | Set the model ID |
| Fully replace the managed system prompt | Set systemPrompt |
| Pin the context management strategy | Set truncation |
| Fix the toolset | List each required tool in allowedTools |
How do I get started?
AWS's demo uses the AgentCore CLI to create a "Research Coordinator" harness that runs research, synthesis and review via subagents in a single invocation. The install command from AWS is below. The CLI uses default AWS credentials and has no --profile flag; set the AWS_PROFILE environment variable if needed.
npm install -g @aws/agentcoreAWS says a full research report run takes about 10 to 15 minutes. Deleting a harness also deletes its managed memory unless you pass --delete-managed-memory false; the execution role created by the CLI stays in your account and you must delete it yourself in the IAM console. AWS also provides a Refund Desk example in the AgentCore samples repository on GitHub, with a six-step walkthrough demonstrating the four events and Lambda, SNS and EventBridge targets.
Frequently asked questions
Do I need to redeploy my existing harness?
According to AWS, no. The upgrade applies automatically to new and existing harnesses, with no customer action and no new resource types or integrations to adopt.
How do I stop the defaults from changing my agent's behavior?
AWS says explicit settings take precedence: set a model ID, systemPrompt or truncation, or list tools one by one in allowedTools, to keep behavior fixed.
Can an SNS or EventBridge hook block a tool call?
No. AWS says only Lambda hooks can change the agent loop; SNS and EventBridge are for notifications only.
Does prompt caching apply to all models?
No. AWS says automatic prompt caching initially targets Anthropic models on Amazon Bedrock and applies to supported models using Converse on Amazon Bedrock.
Can I treat the 34% cost reduction as a guarantee?
No. The figure is AWS's own claim, based on its comparison with other leading harnesses, so treat it as a vendor figure rather than a promised saving.
Browse the latest news in this topic
Found this guide helpful? Share it with someone who needs it
Claude Code 2.1.329: Code Mode Works With Telemetry, /tasks Lists Scheduled Tasks, macOS Computer Use Gets Stricter Safety Rules
Anthropic's changelog lists Claude Code 2.1.329, covering Code Mode behavior changes, new diagnostics in /tasks and /mcp, /code-review changes, and new safety rules for macOS computer use.
Kimi Code CLI v2.7.0 Adds a Native Amazon Bedrock Provider: Use Your AWS Credentials Instead of a Separate API Key
AWS Rounds Up 13 SageMaker Inference Features: Managed NVIDIA Dynamo on HyperPod, MIG GPU Partitioning, Cross-Region Routing
BMW Group Shares How It Governs AI Agents: 12,500+ Developers, Three Autonomy Levels and a Four-Stage CI/CD Gate
Sources
Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.