Code review emerges as primary bottleneck in agentic development
High-velocity code generation has made manual review a critical constraint, prompting a move toward auto-approvable changes. Developers are exploring architectural designs that allow agents to verify code without human line-by-line inspection.
Code review has emerged as a significant bottleneck in agentic development, with developers potentially generating more than 30 pull requests a day. This surge in generation velocity has forced the industry to seek a middle ground where a percentage of code is auto-approved based on architectural design, while humans focus on intent and high-level review. Eric discussed the need for 'auto-approvable' code changes based on thoughtful agent-readiness design and verifiability, citing posts from Cloudflare, Amplitude, and Augment on designing AI code review systems.
The discussion also addressed the challenge of maintaining specifications, with Eric noting that pull request descriptions often serve as the de facto spec. He referenced Basis's approach to version-controlling specs and canonical code states. Eric proposed a 'program database' concept to replace Git, aiming to capture traces, specs, and skills with higher concurrency, arguing that Git's mutexes are a limitation. He suggested that many B2B software companies will evolve into coding agent companies by closing the loop on user insights. The shift represents a move away from line-by-line inspection toward systems that can verify correctness autonomously.