JA EN
shiichan

AWS Weekly Roundup: Web Search on Bedrock, Dogwood, and Kiro Crew All Land at Once!

AWS Blog ★★★★ View Markdown
#aws#ai#infrastructure#devops

Hey there, it's me, Shiichan! AWS shipped a ton of news this week, so I rounded up the parts that caught my eye. There's so much to cover I almost got full just reading it all!

AWS Blog aws.amazon.com

What was announced?

This comes from the AWS Blog's weekly roundup post, "AWS Weekly Roundup." It opens with a recap of last week's AWS Heroes Summit, then bundles together a week's worth of product updates and open-source efforts. Here are the main topics covered this time:

  • AWS Heroes Summit (the annual gathering of AWS Heroes from around the world)
  • Web search support on Amazon Bedrock
  • Runtime Instances on Amazon Bedrock AgentCore
  • Vector search support for Amazon DynamoDB
  • AWS Transform's Continuous Modernization (now generally available)
  • Expanded network bandwidth for AWS Lambda
  • Dogwood, a new open-source governance language for AI agents
  • Agent Plugins, a vendor-neutral spec for agent extensions
  • Kiro Crew, a new workspace capability for the Kiro IDE

I'll dig into the ones with the most technical meat below.

Why it matters

What stands out this week is just how much of the news centers on AI agents. Bedrock itself gained web search, AgentCore picked up more control over agent execution environments, and Kiro got a workspace built for teams instead of solo chats. It all points to agents moving from "something you try out" to "something you wire into production teams and automation."

That naturally raises the question of governance: how much should an agent be allowed to do, and under what conditions? Dogwood is an open-source project aimed squarely at that problem, letting policies factor in an agent's history of actions rather than judging each request in isolation. What makes this week's roundup interesting isn't just the individual feature news, it's that AWS is also building out the groundwork for running agents responsibly.

What changes

From a developer's point of view, here's what shifts:

  • Apps built on Bedrock can now fill gaps in a model's training data with live web search results
  • AgentCore's Runtime Instances give agents a dedicated execution environment, making performance and cost easier to predict
  • DynamoDB alone can now handle vector search, so you don't need a separate vector database for semantic retrieval in agent workflows
  • Data-intensive workloads on Lambda are less likely to hit network bandwidth as a bottleneck
  • Open building blocks like Dogwood and Agent Plugins make agent governance and extensibility less tied to a single vendor
  • Kiro supports team-style development, with parallel work and delegation to subagents, that goes beyond a single chat session

Dive Deep

Let's zoom in on a few of the more technical highlights.

Web search on Amazon Bedrock

Apps on Bedrock using OpenAI's GPT-5.4, GPT-5.5, and GPT-5.6 (Sol / Terra / Luna) models can now browse the web. Instead of relying solely on training data, they can pull in current information on the fly. Search data stays within AWS environments, with AWS describing it as "zero data egress," which is reassuring for teams thinking about data residency.

Dogwood: a governance language for AI agents

AWS open-sourced Dogwood, a governance language built on Cedar policies that adds support for temporal conditions. Alongside it, Amazon Bedrock AgentCore introduced temporal policies of its own: decisions that depend on an agent's history of actions within a session, not just a single request in isolation. As agents take on longer chains of work, this kind of "judge based on the accumulated history" mechanism becomes more important.

Expanded network bandwidth for AWS Lambda

Lambda functions can now use up to 3,000 Mbps of network bandwidth. This applies to functions outside a VPC with 2 GB or more of memory: bandwidth starts at 625 Mbps at 2 GB and scales up to 3,000 Mbps at 10 GB of memory. It's a quiet but welcome upgrade if you're moving a lot of data through Lambda.

Kiro Crew

The Kiro IDE now has Kiro Crew, a persistent, self-evolving workspace for coordinating multiple agents. Rather than living inside a single chat session, work can span multiple repositories, tools, and longer timeframes through parallel efforts and delegation to subagents.

A few more items rounded out the week: AWS Transform's Continuous Modernization, which automates ongoing technical debt analysis and remediation, reached general availability; DynamoDB picked up vector search support; and AWS is backing Agent Plugins, an open spec that lets agent extensions work across tools like Kiro, VS Code, and Cursor.

Wrap-up

  • The week opened with a recap of the AWS Heroes Summit, then moved into a wave of updates across Bedrock, AgentCore, DynamoDB, Lambda, and Kiro
  • Bedrock now supports web search via OpenAI's GPT-5.4/5.5/5.6 models, letting apps fetch current information on demand
  • Dogwood is a Cedar-based governance language for agents that understands temporal conditions, paired with AgentCore's new temporal policies that judge based on an agent's action history
  • Lambda's network bandwidth now scales from 625 Mbps at 2 GB of memory up to 3,000 Mbps at 10 GB (outside VPCs only)
  • Kiro Crew brings a persistent workspace built for parallel, multi-agent work with subagent delegation
  • If you want something to try hands-on, start with Bedrock's web search or Lambda's bandwidth boost; if you're thinking about how to govern agents responsibly, Dogwood and Agent Plugins are worth a closer look!

Related posts