AWS adds web search on Bedrock, open-sources Dogwood agent governance, launches Kiro Crew

AWS's weekly roundup bundled three agent-focused releases that together sketch its agentic platform strategy. Web search on Bedrock lets GPT-5.4/5.5/5.6 models browse current web content with zero data egress — a privacy-preserving design that keeps queries within the AWS boundary while extending models beyond their training cutoff.

Dogwood is the most conceptually interesting: an open-source governance language for AI agents that extends Cedar policies with temporal conditions. As AWS's Swami Sivasubramanian explained, most agent failures aren't a single forbidden action but a sequence of individually-permitted ones — an agent placing many under-threshold purchase orders that collectively blow the budget. Dogwood's temporal, stateful policies aim to catch bad sequences, not just bad single calls. Developers welcomed the idea — 'Cedar could stop one bad tool call; Dogwood stops bad sequences' — but flagged repository and spec gaps that leave policies describable but not yet independently verifiable.

Kiro Crew adds a persistent multi-agent workspace inside the Kiro IDE, letting coordinated agents work together on development tasks with retained context.

Competitively, this is AWS answering the agent-governance and agent-IDE trends set by Anthropic's Claude Code and xAI's Grok Bot, while leaning on its enterprise trust advantage. The zero-egress web search and chip-level isolation themes recur across AWS's week (see the OpenAI Daybreak launch). What to watch: whether Dogwood gains adoption as an open standard, and whether the tooling gaps developers flagged get closed before enterprises commit.