Operationalizing AI-Native Development at Maven Clinic
AI Engineergo watch the original →
the gist
Maven Clinic transitioned to an AI-native workflow by shifting from long-term planning to two-week sprints, implementing self-certified code reviews for high-volume AI-generated code, and using multi-model consensus to handle non-deterministic reliability requirements.
Adapting Planning and Hiring for AI Velocity
Maven Clinic shifted its development philosophy because the cost of implementation has plummeted while the cost of debating requirements remains high. The company now treats one-year roadmaps as directional guidance and focuses on two-to-four-week sprints for concrete execution. This approach renders three-to-six-month plans effectively unplannable, as the rapid evolution of AI model capabilities makes long-term technical commitments obsolete. Hiring now prioritizes engineers who can independently solve problems using AI tools, effectively blurring the lines between product management and engineering roles.
Managing High-Volume Code Review
With AI tools enabling engineers to write thousands of lines of code daily, traditional manual review processes have become a bottleneck. Maven Clinic implemented the following changes to maintain velocity without sacrificing quality:
- Engineers self-certify whether a pull request requires a second reader, holding them accountable for the quality of unreviewed code.
- Pull requests are capped at 500 lines to ensure reviewers can provide meaningful feedback.
- Large features are broken into stacked pull requests to maintain a continuous development flow.
- The team explicitly forbids "rubber stamping" pull requests, as it creates false confidence in the system.
Reliability and Non-Deterministic Systems
Because generative AI models are non-deterministic, the team categorizes failures into tolerable and intolerable buckets. For low-stakes tasks like scheduling, occasional failures are accepted as a cost of doing business. For high-stakes tasks like reimbursement claims, the company enforces a zero-tolerance policy by requiring multiple models to process the same input and only proceeding if the outputs agree. To validate these systems, integration tests are run repeatedly rather than once to account for variance, and the company maintains a rigorous evaluation loop using automated rubrics combined with manual spot-checks of 20% of new feature interactions.