FurtherAI co-founder Sagar Gondalia recently shared a detailed breakdown of how his team transformed their engineering organization into an AI-native entity over the last nine months. The results are striking: since January, the number of code changes each engineer merges per day has grown approximately 5x, while the revert rate remained stubbornly low at under half a percent. This wasn't achieved by simply handing out Copilot licenses; it required a systematic overhaul of every stage in the shipping pipeline.

Rebuilding the Shipping Process

The transformation was incremental and targeted specific bottlenecks month by month. By March, agents were writing code with engineers steering the direction, reaching nearly 100% adoption. In April, test coverage jumped from 15% to 91% as agents began writing tests alongside code. By May, AI reviews covered 99% of changes before human intervention, allowing reviewers to focus solely on high-judgment areas. The CI pipeline was overhauled in June to run 20x more tests in the same 17-minute window by parallelizing execution and only running affected tests.

Automating Triage and Merging

July marked the integration of Claude Tag for bug intake, which now handles 60% of ticket filing by identifying duplicates and assigning owners directly from Slack reports. In August, the team introduced 21 custom Python rules that convert tribal knowledge into automated checks, enabling low-risk changes to merge without human reviewβ€”a process that now accounts for up to a third of all merges. By September, they launched 'code channels,' allowing non-engineers to describe changes in Slack and have agents open PRs, a method heavily utilized by their design team, who merged about 50 changes since August.

The New Bottleneck: Planning and Review

Despite the speed gains, a new constraint emerged: decision-making. Gondalia notes that while building code became 3.5x faster, planning and review phases remained static, causing them to jump from 40% to 70% of the total feature timeline. The traditional model of flowing decisions through engineering managers and PMs is no longer viable when engineers can ship at this velocity. FurtherAI is now pushing more ownership to individual engineers, redefining seniority by the scope of problems they can own end-to-end rather than just their coding prowess.

Key Takeaways

  • Velocity is a system problem: Optimizing code generation alone creates bottlenecks in testing, CI, and review.
  • Non-engineers are now contributors: Designers and customer-facing staff are shipping UI tweaks and design system changes via Slack agents.
  • Managerial roles are shifting: Engineering managers must become more product-savvy and deeply involved in the user experience, not just code review.
  • Tribal knowledge is code: Converting implicit rules into automated Python checks allows AI agents to self-correct and merge low-risk changes independently.

The Bottom Line

This isn't just about faster coding; it's about dismantling the traditional approval chains that stifle velocity. If your org is still gating every PR behind a human manager while agents write the code, you're already obsolete.