AWS AgentCore’s Strands upgrade can change omitted agent settings
Strands-backed AgentCore harnesses began rolling out on Nov. 11, adding evolving model and tool defaults plus four Lambda checkpoints that can deny actions before they execute.

Why it matters
- Strands-backed AgentCore harnesses began rolling out on November 11, 2026; AWS announced the upgrade on November 19.
- Strands was released earlier in 2026 under the Apache 2.0 license.
- Lambda lifecycle hooks can deny actions at four checkpoints; SNS and EventBridge hooks only notify.
- A support-agent example denies `issue_refund` calls above a $500 auto-approval limit.
- Four explicit controls—model ID, `systemPrompt`, `truncation` and `allowedTools`—determine whether defaults can evolve.
By James Calloway
AWS's Amazon Bedrock AgentCore has four Lambda lifecycle checkpoints (fixed policy checks in the agent loop) that can deny tool calls, including an issue_refund above a $500 auto-approval limit. The Strands-backed upgrade began rolling out to new and existing harnesses on November 11, 2026, and AWS announced it on November 19. It adds managed defaults, built-in tools and those hooks without requiring a new resource type or integration.
That design puts a stability tradeoff at the center of the release. Workloads that require fixed behavior must explicitly preserve their model, instructions, context policy and tool surface (the tools exposed to an agent), because AWS can revise recommendations for anything left unspecified.
What changes in AgentCore?
AWS released Strands earlier in 2026 under an Apache 2.0 license. The open-source project assembles an agent foundation covering instructions, tools and context management. AgentCore operates that foundation as a managed AWS service, with each session running in a kernel-isolated microVM (a lightweight virtual machine with a dedicated kernel).
AWS manages compute, scaling, high availability and the infrastructure that executes tools. Customers retain control over models, instructions, tools, permissions and lifecycle policies.
Takeaway: AWS owns the operating layer, while customers retain control of the agent's policy boundaries.
What do managed defaults control?
AWS's announcement says agents using managed defaults receive a “recommended model that can change as new models are released.” The service can also silently update its task-oriented system prompt as AWS revises that recommendation.
That prompt directs agents to gather relevant context, use appropriate tools, adapt to failures or denied actions, and verify their work before reporting completion. AWS can revise these unspecified settings without customer action.
The convenience comes with a caveat: the announcement gives no schedule for revisions and no comparison of old and new prompts or models. Teams cannot determine from the release alone how much behavior may shift.
Takeaway: Managed defaults exchange exact reproducibility for the flexibility to follow AWS recommendations.
Where can lifecycle hooks stop an agent?
Only AWS Lambda hooks can change the agent loop, and AWS placed four deterministic checkpoints within it. Lambda functions run synchronously and return an allow or deny decision, optionally with a reason.
before_invocationruns before the agent processes the request. A denial stops the invocation and returnshook_stopped.before_tool_callruns after the model proposes a tool call but before execution. A denial skips that call while the agent loop continues.after_tool_callruns once a tool result exists. A denial retains the result and stops the harness before further inference.after_invocationruns after the invocation completes. A denial reports the decision, but AWS says streamed output cannot be retracted.
Amazon SNS and Amazon EventBridge hooks serve a narrower role. They send notifications without blocking execution, so they cannot provide the same denial behavior as Lambda.
Takeaway: Lambda hooks can stop or divert an agent, while notification hooks only observe it.
What happens when a hook fails?
If a Lambda function times out or returns an error, AWS applies the configured failureMode. The default is deny, so an unhandled failure produces a denial.
Each decision appears as a hookEvent entry in the InvokeHarness response stream. That gives operators a record in the response path, although the announcement does not describe separate audit storage or retention.
Takeaway: A failed hook blocks the governed path by default, and its decision remains visible in the response stream.
How can a hook stop a $500 refund?
AWS's support-agent example shows the intended policy boundary. A before_tool_call hook denies an issue_refund call above a $500 auto-approval limit.
The harness skips the refund and continues the agent loop, allowing the agent to report an escalation. Tool-call hooks also govern subagent calls, which prevents delegation from bypassing the same policy. Invocation-level hooks run once for the overall invocation.
The example demonstrates control flow, but AWS provides no added-latency figure or failure-rate data. It also does not quantify the operational cost of running Lambda checks at each applicable checkpoint.
Takeaway: The hook path blocks an action before execution and gives the agent a route to escalate.
How does AWS manage growing context?
Automatic prompt caching is initially available for supported Anthropic models through Amazon Bedrock Converse. It covers the system prompt, tool definitions and conversation history.
AWS says repeated calls with the same prompt and tool surface can use lower-priced cache-read tokens, meaning tokens already processed and stored for reuse. The company says this can reduce input-processing latency and cost.
The announcement provides no cache-hit rate, price comparison or latency benchmark. The claimed savings therefore remain directional rather than quantified.
Proactive context management works differently. After context use crosses a configurable threshold, it summarizes older conversation turns while retaining recent context. Separately, context offloading moves large tool results out of the model's immediate context. The agent receives a preview and can retrieve the complete result when needed.
Neither mechanism specifies summary fidelity or an offload-size threshold. Those omissions matter for agents whose decisions depend on exact earlier text or large outputs.
Takeaway: Caching targets recurring input costs, while summarization and offloading trade immediate context for reachability.
Which tools ship with the harness?
AWS added several built-in tools under the Strands-backed harness. todo_write preserves a structured plan across invocations in the same session, including after summarization.
The subagent tool assigns a self-contained task to a fresh worker inside the current invocation. That worker shares the invocation's limits and tool-call hooks, and it does not require a separately managed harness resource.
web_fetch retrieves webpage content. Separate read, write and edit tools let customers grant filesystem permissions individually rather than exposing one undifferentiated capability.
That granularity supports narrower controls, although the announcement does not state whether its built-in tool list is exhaustive. The release also gives no detail about how web_fetch treats untrusted page content.
Takeaway: The new tools add planning, delegation and file access while retaining granular permissions and tool-call controls.
Which settings keep behavior stable?
The Strands-backed experience began rolling out to new and existing harnesses on November 11, 2026, and requires no customer action. Explicitly supplied configurations remain unchanged.
Existing Model Context Protocol (MCP) servers, AgentCore Gateway, Browser, Code Interpreter, memory and customer-provided containers continue to work through their current interfaces. AWS did not describe a separate migration for those integrations.
Configuration determines how much behavior can change later:
- A model ID keeps a specific model fixed.
systemPromptreplaces the managed instructions.truncationretains a chosen context-management policy.- Omitting
allowedTools, or setting it to*, accepts a recommended tool set that AWS says can “gain capabilities over time.” - Listing individual tool selectors maintains a fixed security or capability boundary.
AWS can revise recommendations wherever fields remain omitted. The announcement does not promise that future recommendations will preserve existing security boundaries, so explicit tool selection carries more weight than a broad wildcard.
Takeaway: Pinning model, prompt, truncation and tools turns the service's evolving defaults into a controlled configuration.
As AWS revises its recommendations, the distinction between omitted and explicit settings will decide whether an agent keeps the same operating profile. Customers needing stable behavior can retain that profile through fixed fields, but the default-deny timeout behavior also makes failure-path testing a practical requirement.
That is the release's practical consequence: more managed convenience now, with a larger burden on teams that need predictable behavior to document and enforce their own boundaries.
Source: unite.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- AWS AgentCore Runtime Gets 8-vCPU Sessions and Filesystem Mounts
- AWS Open-Sources UAEF, a Framework for Evaluating AI Agents Across Rivals
- AWS Turns Step Functions Into Agent Tools as GPT-6.1 Terra Hits Bedrock
- AWS Opens Direct Web Search API at $7 per 1,000 Queries
- OpenAI Models Are Coming to Amazon Bedrock via a Stateful Agent Runtime