AI 热点

AWS 为 Amazon Bedrock AgentCore harness 引入…

Unite.AI更新于 11月20日 05:07

AWS Adds Strands Defaults and Lifecycle Hooks to AgentCore Harness

AI 导读

AWS 升级 Amazon Bedrock AgentCore harness,采用开源 Strands harness,并加入可演进的托管默认配置、内置工具和确定性生命周期钩子。钩子可在调用及工具执行前后通过 Lambda 允许或拒绝操作,相关限制同样覆盖子智能体调用。新体验自 2026年11月11日开始向新旧 harness 推送,无需客户操作,现有显式配置保持不变。

推荐理由

AI 评分 70

新钩子为智能体调用工具和子智能体增加确定性控制,显式配置也成为保持模型与工具边界稳定的关键。

主题

标签

正文

Amazon Web Services announced on November 19, 2026 that it is upgrading Amazon Bedrock AgentCore harness to use the open-source Strands harness, adding evolving managed defaults, additional built-in tools and deterministic lifecycle hooks without requiring a new resource type or integration.

Strands, released earlier in 2026 under an Apache 2.0 license, provides an assembled 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. AWS manages compute, scaling, high availability and tool-execution infrastructure, while customers retain control over models, instructions, tools, permissions and lifecycle policies.

Agents using managed defaults receive a recommended model that can change as new models are released, along with a task-oriented system prompt that AWS also updates over time. The prompt encourages agents to gather relevant context, use appropriate tools, adapt to failures or denied actions, and verify their work before reporting completion.

Automatic prompt caching is initially available for supported Anthropic models through Amazon Bedrock Converse, covering 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, reducing input-processing latency and cost.

Proactive context management summarizes older conversation turns after context usage reaches a configurable threshold while retaining recent context. Separately, context offloading moves large tool results out of the model’s immediate context, leaving the agent a preview and access to the complete result when needed.

The built-in tools include todo_write, which preserves a structured plan across invocations in the same session, including after summarization. A subagent tool assigns a self-contained task to a fresh worker inside the invocation, sharing its limits and tool-call hooks without creating a separately managed harness resource. web_fetch retrieves webpage content, while separate read, write and edit tools allow filesystem permissions to be granted individually.

Lifecycle hooks insert deterministic controls at four points in the agent loop. A Lambda denial has a different effect at each checkpoint:

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; streamed output cannot be retracted

Only Lambda hooks can change the agent loop. They run synchronously and return an allow or deny decision with an optional reason; Amazon SNS and Amazon EventBridge hooks provide notifications without blocking execution. If a Lambda function times out or returns an error, the configured failureMode applies, defaulting to deny. Decisions appear as hookEvent entries in the InvokeHarness response stream.

In AWS’s support-agent example, a before_tool_call hook denies an issue_refund call above a $500 auto-approval limit. The harness skips the refund and lets the agent continue so it can report an escalation. Tool-call hooks also govern subagent calls, preventing delegation from bypassing the policy, while invocation-level hooks run once for the overall invocation.

Rollout Preserves Explicit Configurations

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 customer Model Context Protocol servers, AgentCore Gateway, Browser, Code Interpreter, memory and customer-provided containers continue to work through current interfaces.

Configuration settings determine how much of an agent’s behavior can change later. A model ID keeps a specific model, systemPrompt replaces the managed instructions, and truncation retains a chosen context-management policy. Omitting allowedTools or setting it to "*" accepts a recommended tool set that can gain capabilities over time; listing individual selectors maintains a fixed security or capability boundary.

The managed service can therefore update unspecified parts of an agent as AWS revises its recommendations, making explicit configuration the control point for workloads that require stable behavior and a defined tool surface.

来源

Unite.AI