Sashank Gondala, building at FurtherAI, has published a detailed operational log of what happens when you give a small engineering team coding agents and then systematically remove every bottleneck that surfaces downstream. The results are specific: code changes per engineer per day up 5×, revert rate under 0.5%, test coverage jumping from 15% to 91%, and CI running 20× more tests in the same 17-minute window. Up to a third of changes now merge without a human touching them. The sequencing matters more than the individual numbers. FurtherAI didn't just hand engineers Copilot and call it done. They walked the entire shipping pipeline month by month: agents write code (March), agents write tests (April), AI pre-reviews PRs (May), CI parallelized (June), bug intake automated via Claude (July), tribal knowledge encoded as automated rules (August), low-risk auto-merge enabled (August), and Slack-to-PR channels opened for non-engineers (September). Each optimization exposed the next chokepoint. The pattern is factory-floor Lean applied to software delivery. The cost was real. Through the summer, customer-facing changes per engineer sat between 17 and 30 per month while the team invested in tooling and test infrastructure. September jumped to 45; October is pacing at 48. The J-curve of investment before throughput is honest and rarely disclosed this clearly by companies selling an AI narrative. The more interesting admission is what broke next: decision-making. Gondala's simplified model — a feature that once took ten days (three planning, six building, one review) now takes 5.7 days because building got 3.5× faster but planning didn't change at all. Planning and review went from 40% to 70% of elapsed time. The bottleneck migrated from hands to heads. Gondala's organizational response is to push ownership downward. His redefined seniority ladder — junior engineers steer agents and detect errors, senior engineers plan and review their own work, staff engineers own ambiguous problems end-to-end and define their own success metrics — is essentially a compression of the traditional PM-EM-designer triad into the engineer role. Non-engineers shipping code (a designer with 50 merged changes since August) further blurs the boundary. The tension Gondala doesn't fully resolve is whether this model scales beyond a small, high-trust startup. Encoding tribal knowledge as Python checks works when there are 21 rules. At 2,100 rules with competing stakeholders, the governance problem re-emerges. Auto-merging a third of changes with sub-0.5% revert rates is impressive now, but the tail-risk surface grows nonlinearly with codebase complexity and team size. The approach is genuinely generative at current scale but carries latent fragility. This is an unusually transparent operational post from a company that is also hiring for the roles it describes. The self-interest is visible but the specificity of the numbers — revert rates, test percentages, monthly change counts, timeline of each optimization — makes it more useful than the typical "AI transformed our org" thought piece. The real question it poses to every engineering leader: if your code-writing bottleneck disappears tomorrow, what breaks next?
