How to build an AI-Native Health Company — Dan Feng, Maven Clinic
Maven’s agent-heavy workflow shortens planning, caps reviewable PR size, and tests nondeterministic features repeatedly. The bottleneck shifts from implementation to decisions and verification.
Maven describes replacing long planning cycles with **2–4-week delivery windows**, **one-to-two-page specs**, and reviewable PRs capped at **500 lines**. Its LLM integration tests target a sustained **90% pass rate**, not a single green run.
Keep agent-generated changes small, make short-term requirements concrete, and spend engineering attention on architecture, evaluation, and consequential failure paths. Low-risk changes may justify self-review, but ownership stays with the author.
Maven describes replacing long planning cycles with **2–4-week delivery windows**, **one-to-two-page specs**, and reviewable PRs capped at **500 lines**. Its LLM integration tests target a sustained **90% pass rate**, not a single green run. Keep agent-generated changes small, make short-term requirements concrete, and spend engineering attention on architecture, evaluation, and consequential failure paths. Low-risk changes may justify self-review, but ownership stays with the author. These are one healthcare company’s operating rules, not established defaults. Maven still values human review, says automated review is not yet sufficient, and has not completed its goal of automating the full delivery lifecycle.
This turns bounded agent engineering into specific operating constraints: short delivery windows, compact specs, small PRs, and repeated integration-test performance. It confirms that higher automation does not remove authorship or human review; engineering effort shifts toward architecture, evaluation, and costly failure paths. The numbers are local practices, not general thresholds.