When agents make coding 3.5 times faster, a 10-day feature still takes about 5.7 days, and planning and review eat 70% of it. How to shrink those two stages.
What OpenAI’s repository problem reveals about testing, review, and decisions—and how to identify the constraints in your own work.
TL;DR
- OpenAI's coding agents accelerated development, but their infrastructure struggled to keep up.
- Pando was built by OpenAI to solve repository-related issues, leading to a nearly 80% drop in median push latency.
- Despite agent-driven speedups, delays shifted to testing, review, and decision-making stages.
- FurtherAI achieved a fivefold increase in merged changes per engineer, highlighting the need to optimize downstream processes.
- Automated checks and review rules are being implemented for agents to fix issues before human review.
- The article discusses strategies for identifying and addressing development constraints, including the concept of throwaway prototypes and constraint reviews.