Baz has launched a Spec Review agent that uses large language models hosted on Amazon Bedrock and the Amazon Bedrock AgentCore runtime to automate verification that code changes fulfill product and design requirements. The tool brings product intent and behavioral validation into code review, aiming to close the gap between diff-focused reviews and the runtime, visual, and design checks that traditionally occurred manually late in the development cycle. This change is meant to surface misalignments earlier in the pull request lifecycle and reduce manual QA overhead.
The implementation runs as a multi — stage pipeline that can be triggered by a GitHub webhook or invoked manually. At the start of a review, the agent concurrently queries Figma via MCP and Jira over REST to collect design files, visual specifications, and functional artifacts. Those artifacts are then passed to a Specification Subagent to extract discrete requirements and to a set of Implementation Subagents that execute validation work.
Baz spawns isolated subagent workers — one per extracted requirement — that combine static repository checks with dynamic runtime validation. The Specification Subagent, powered by Bedrock LLMs, decomposes incoming artifacts into discrete visual requirements (spacing, colors, hierarchy) and functional requirements (acceptance criteria, user-story intent), producing the individual tasks each subagent will verify. Each Implementation Subagent uses AgentCore’s Browser tool to perform live validation inside temporary Preview Environments. Subagents execute DOM inspection, event simulation, and visual testing to compare deployed UIs against Figma visual specifications and Jira acceptance criteria instead of relying solely on static code analysis. That live testing approach is intended to catch UI regressions and behavioral mismatches that code diffs alone can miss.
Traffic for live checks is routed through an Application Load Balancer (ALB) and Network Load Balancer (NLB) into an Amazon EKS cluster, where the Baz Platform orchestrates the agentic review and isolates per — PR environments. This architecture supports concurrent, ephemeral validation runs tied directly to the pull request under review. For engineering teams and builders, Baz says the pattern delivers end-to-end verification inside the pull request workflow and reduces manual QA effort that often required clicking through preview environments and relying on undocumented testing knowledge. Running checks per PR in isolated environments helps surface misalignments and potential regressions earlier and more consistently.
Baz and AWS present the design as a repeatable architecture: aggregate design and product artifacts, decompose requirements into subagents, use AgentCore for live UI verification, and rely on Bedrock for scalable, secure inference. The post does not publish operational metrics but frames the agent and pattern as a blueprint builders can adapt to integrate design tools, issue trackers, and runtime checks into CI workflows.
Sources
Replies (0)
No replies in this topic yet.
