Vercel Eve: AI Revolutionizes Code Review in Enterprises

An AI Agent for Code Review: A Revolution in 30 Minutes

Claire has developed an artificial intelligence agent capable of reviewing code merge requests, assessing their risk level, automatically approving those deemed safe, and flagging the more questionable ones on Slack. This system was designed in a single session with Codex, using Vercel Eve, a tool that could be a game changer in managing AI-generated code.

Proven Efficiency with AI

The notion that every AI-generated merge request requires human verification is being challenged. While this may seem risky, companies like Intercom have demonstrated that AI systems can handle these tasks at scale. Merge requests approved by AI progress five times faster than those reviewed by humans, with a lower return rate. This means that AI-validated code is not only deployed more quickly but also requires fewer subsequent corrections.

A Well-Defined Risk Model

To differentiate changes that require human intervention from those that do not, a risk model has been established. This model evaluates each merge request based on six criteria: the scope of the change, its potential impact, the ease of rollback, data and security implications, operational impact, and the completion of testing and continuous integration. Requests rated below 24 points are considered low risk and are automatically approved, while those above 64 points are sent to a human for review. This system transforms a subjective judgment into a systematic process.

Vercel Eve: A Fast and Efficient Solution

Vercel Eve is touted as the fastest tool for deploying an internal AI agent on Slack and GitHub without spending weeks building the necessary infrastructure. It handles tedious technical aspects, such as connectors, refresh tokens, sandboxing, and routing through channels, allowing developers to focus on writing instructions and Markdown skills rather than managing OAuth flows.

Simplified Deployment with Codex

Creating and deploying an internal agent can now be accomplished in a single Codex session. The initiation of the process begins with a simple request: build a GitHub bot that evaluates merge requests based on their risk and automatically approves those that are low risk. The rest of the process involves adjustments and refinements, without requiring a complex initial specification.

The Importance of User Interface

Using the browser significantly simplifies the setup of agents, eliminating the complexity that often makes their implementation difficult. Creating a Slack bot and a GitHub application manually typically involves navigating through numerous permission screens and managing tokens. Codex has streamlined this process, reducing the time required from several hours to just a few minutes.

Compliance and Automation: A Possible Balance

SOC 2 compliance and automatic approval of merge requests are not incompatible. The key is to make the process transparent: the risk model must be integrated into the company's code review and security policies, every decision must be documented, and the audit trail must be easy to query and defend. The security team plays a crucial role in helping to design the appropriate framework, rather than blocking automation out of fear of the unknown.

Maintaining Human Accountability

The operational architecture is as crucial as the technology itself. The Merge Mommy agent, for example, does not directly merge requests. It publishes a signal in GitHub and sends a message on Slack with the risk score, indicating that the request is ready to be approved and merged. This process preserves human accountability while reducing the cognitive load associated with reviewing routine changes.

Trust in AI Agents

Regular evaluations are essential to maintain trust in internal agents after the novelty effect wears off. Intercom, for example, logs every merge request review performed by its agent, and then an engineer assesses the accuracy of the score and recommendation. This discipline is the same as that applied to AI products intended for customers. Internal agents, while less visible, require similar protection against regressions.

Simplicity and Efficiency of the Eve Agent

Creating an agent like Eve does not require complex construction. The instructions for Merge Mommy fit on a single page: a few paragraphs, some bullet points, and a short skills file. The main skill lies in the ability to clearly and precisely explain what the agent should accomplish.

Claude Code: AI at the Service of Businesses

Grace Clarke, an AI educator and former marketing consultant, has reinvented her business around Claude. She highlights three key skills of Claude that support her business: a timely client pipeline, a proposal generator, and a voice guide that teaches Claude her way of thinking. She has also replaced Gmail with a Claude-powered inbox, emphasizing the importance of intention engineering over prompt engineering.

Building Without Technical Skills

You don’t need to be a technical expert to create a business workflow with Claude. Grace started by opening Claude Code and expressing her frustrations: too many emails, a lack of warm client communication, and 20 hours of weekly administration. Claude transformed these complaints into an operating system for her business.

The Importance of Intention Engineering

For Grace, intention engineering is more crucial than prompt engineering. She is not trying to create the perfect prompt but rather to explain the problem, describe the desired outcome, and ask Claude to propose a solution. Her philosophy is that the burden of the next step should not always rest on the person. With enough context, Claude should be able to propose solid ideas.

A Personalized Voice Guide

Grace's voice guide is not just a writing guide but a file that captures her way of thinking. It contains her decisions, her opinions on teaching, and even the types of LinkedIn posts she avoids. Whenever she encounters something that bothers her, she sends a voice note to Claude to update the guide.

An Optimized Workflow

By moving her inbox workflow into Claude, Grace ensures that each interaction becomes useful context for the next. Unlike Gmail, where context remains buried, the AI can leverage this information to improve.

Integrating Claude Code and Cowork

Grace has developed an elegant method for transitioning from Claude Code to Cowork. She uses Claude Code for proactive tasks and then transfers the context into Cowork for a more visual environment. This allows for transporting context without starting from scratch.

Overcoming Barriers to AI Adoption

The biggest barrier to AI adoption may be muscle memory rather than fear. Grace encourages her students to integrate AI into their routine, for example, by sending Slack reminders to capture their work and share it with Claude.

The Importance of Skills Files

Skills files are an underestimated idea in agentic AI. Grace manages her business around three main files: a pipeline operator, a proposal generator, and a voice guide. Each file is a set of instructions that Claude can reuse. She teaches her students to build them as they would train a new employee.

Demonstrating AI in Action

To convince clients of the effectiveness of AI, Grace offers an interactive onboarding experience to her new clients. This demonstration allows them to concretely understand the benefits of this new approach even before she explains it.