Code review was always manual and inefficient due to the inherent disconnect between code and product. Developers could check whether the code compiled and worked, but not whether it met all functional and design requirements. In the past, QA teams spent hours manually clicking through preview environments to ensure features behaved as expected, and even more time aligning implementations with design intent. This manual validation slowed delivery, introduced inconsistencies, and increased the likelihood of regressions. With the speed of development teams increasing, Baz wanted to automate this missing review layer, bringing together intent, behavior, and implementation into a single review workflow.
This article explains how Baz created his Spec Review agent using Amazon Bedrock and Amazon Bedrock AgentCore. We’ll discuss the architectural decisions, implementation details, and business outcomes achieved by leveraging these AWS services to automate their code review process.
The main problems Baz is trying to solve
Baz is designed to go beyond traditional, differences-only evaluations and validate whether a feature meets the intended product requirements. Early on, Baz found that teams struggled with reviews that focused on syntax rather than behaviors, leaving critical questions like “does it work,” “does it match specs,” “does it behave as expected” to have to be answered manually and late in the process. This gap between code and product intent slowed the team down, created design inconsistencies, and required a heavy reliance on undocumented internal QA knowledge. Baz decided to bridge this gap by creating agents that could evaluate not only the code, but also the experience actually delivered.
Solution overview
The Baz Spec Review agent orchestrates a sophisticated, multi-step validation pipeline: upon trigger (webhook or manual call), it simultaneously queries Figma via MCP and Jira via REST APIs to aggregate comprehensive requirements artifacts spanning technical, product, and design specifications. The system then generates isolated sub-agents (one per requirement) responsible for verifying the requirement. This subagent combines code verification through the source code repository with dynamic runtime validation using the Amazon Bedrock AgentCore navigation tool. The subagent interacts with temporary environments, performing DOM inspection, event simulation, and visual testing to ensure the deployed implementation matches both Figma design specifications and behavioral requirements, providing end-to-end verification throughout the lifecycle from specification to implementation via native AWS orchestration.
The following diagram illustrates the Spec Reviewer architecture, a joint solution from Baz and AWS that enables automated design and product validation within your code review workflow. The entire agent flow is powered by large language models served through Amazon Bedrock, providing scalable and secure AI inference throughout the pipeline. The flow begins when a GitHub webhook triggers on a new pull request, routing traffic through an Application Load Balancer (ALB) and a Network Load Balancer (NLB) to an Amazon EKS cluster. The Baz Platform serves as a central orchestration layer, coordinating the multi-agent review process.
Within the Amazon EKS cluster, Baz’s Spec Review agent breaks down the validation workflow into specialized subagents. The specification subagent, powered by Amazon Bedrock, ingests both visual specifications from Figma and functional specifications from Jira, then decomposes them into discrete requirements: visual requirements (such as spacing, colors, and component hierarchy) and functional requirements (such as acceptance criteria and user story intent).
Implementation subagents are at the heart of this architecture. These Amazon Bedrock-powered agents perform in-depth code analysis against extracted specifications, but what sets them apart is their integration with the Amazon Bedrock AgentCore browser usage functionality. Rather than relying solely on static code analysis, implementation subagents can render the actual implementation in a live preview environment and visually validate that the UI matches the intended Figma designs and that the functionality behaves as specified in Jira. This combination of code understanding and browser-based validation allows Baz to spot discrepancies that traditional code review tools would miss entirely.
A report generator consolidates findings from all subagents into a cohesive review summary. Once the review is complete, results are distributed to the appropriate channels: comments are posted directly to GitHub PR, notifications are sent to Slack for team visibility, and identified issues can be automatically linked to Jira for tracking and resolution.
How Baz implemented Amazon Bedrock AgentCore to address these challenges
Amazon Bedrock AgentCore became the basis for creating an AI code reviewer that can validate actual product behavior. Its secure, isolated, serverless browser sessions allow the Spec Reviewer agent to open preview environments, navigate features, and examine user interface behavior exactly as a user would. By combining the Amazon Bedrock AgentCore runtime to run MCP servers that integrate with ticketing systems, the Amazon Bedrock AgentCore navigation tool with lightweight automation and context modules, Baz Reviewer can compare live behavior and code to ticket and design specifications without requiring browser infrastructure or custom orchestration. Amazon Bedrock AgentCore’s isolation, sandboxing, and observability help Baz scale multiple MCP servers and enable the agent to perform full-stack validation at scale securely and reliably.
Enabling Intelligent Code Review with Amazon Bedrock
Amazon Bedrock powers the reasoning and decision-making layer behind the Spec Reviewer agent, allowing it to interpret requirements, understand design intent, and evaluate the appropriateness of behaviors observed in the browser. Using core models managed by Amazon Bedrock, the agent can synthesize specification context, analyze user interface states, and produce accurate, actionable conclusions about whether a feature meets expectations. Amazon Bedrock provides the reliability, security, and scalability needed for production agent workflows, allowing Baz to offload complex interpretation and validation logic to a high-performance LLM while keeping browser execution isolated in AgentCore. This combination allows the reviewer to bridge the gap between what was planned and what was actually built.
Conclusion
The Baz Agent Spec Review shows how Amazon Bedrock and Amazon Bedrock AgentCore enable organizations to automate product approval workflows that previously required significant manual effort. By leveraging Amazon Bedrock core models for requirements interpretation and decision making, combined with the secure automation capabilities of the AgentCore Browser, Baz has created a solution that validates implementations against specifications throughout the development lifecycle, reducing reported bugs by up to 50% and merge time by 30-70%.
Customers who have adopted Spec Reviewer have seen a significant reduction in manual product validation work, with feature verification performed earlier in the development cycle and occurring automatically during pull requests. Teams report faster reviews, fewer regressions, and greater confidence that changes meet pre-merge requirements.
About the authors
Guy Eizenkot
Guy Eizenkot is the co-founder and CEO of Baz. Previously, Guy was co-founder and VP of Product at Bridgecrew, a company acquired by Palo Alto Networks, where he later led Prisma Cloud’s application security business and helped evolve its application security product line. Prior to Bridgecrew, he served as product director focusing on applied machine learning, cloud security, and large-scale security platforms. Guy is passionate about the intersection of AI and software engineering, developer workflows, and creating products that reshape how engineering teams work. Outside of work, he enjoys playing tennis and squash and spending time with his 3 children.
Nimrod Kor
Nimrod Kor is the co-founder and CTO of Baz, where he leads the company’s AI engineering and architecture efforts focused on transforming how developers review and ship code. Before founding Baz, Nimrod worked on cloud infrastructure, developer tools, and large-scale distributed systems, with a focus on performance and developer experience. Passionate about AI-powered software engineering and open source development, he actively shares technical information and creates tools for modern engineering teams. Outside of work, he is an avid surfer and traveler who spends as much time as possible near the ocean.
Father Atas
Father Atas is a startup solutions architect at Amazon Web Services. He works with startups to help them build and design their cloud solutions, and is passionate about machine learning and container-based solutions. In his free time, Itay enjoys DIY projects and cooking.
Read more about how Baz improved its AI agent code review accuracy using Amazon Bedrock AgentCore here.
“`
