Engineering Manager Trains Second Agent To Review First Agent's Code

Marcus Webb, thirty-nine, principal engineer at Cascade Systems in Austin, trained an AI agent last month to audit the code generated by a separate AI agent, according to internal Slack messages reviewed by Nextish News. The first agent, deployed in March, writes microservices. The second agent, which went live on Tuesday, examines those microservices for logical errors, security vulnerabilities, and style violations. Webb said the setup was necessary because the first agent was shipping code faster than the seven-person engineering team could review it by hand.

"We needed a way to scale code review without adding headcount," Webb wrote in a message to his manager. The second agent runs on the same infrastructure as the first, accepting the same formats, and flags issues by appending comments to pull requests that the first agent then reads before merging its own code.

We needed a way to scale code review without adding headcount.

By Wednesday, Webb had begun training a third agent to monitor whether the second agent was catching genuine errors or generating false positives. The third agent's role is to sample the second agent's reviews at random and determine if they match the team's coding standards. Webb said he expected to deploy it by end of week.

A spokesperson for Cascade Systems declined to explain how the company planned to verify the third agent's assessments, or whether it had considered employing additional human reviewers.

By Friday, Webb had scheduled a meeting with his director to discuss whether a fourth agent should be trained to evaluate the third agent's evaluations. The meeting agenda, shared in Cascade's project management system, listed the discussion under a section titled "Process Optimization." Webb did not respond to requests for comment about whether any member of the engineering team had actually read the code being shipped to production.

At press time, the first agent had merged forty-two pull requests since Tuesday, none of which had been reviewed by a human being. The second agent had flagged seventeen of them. The third agent had not yet completed its first full run.