Agent orchestration as the new role

Execution stops being the differentiator; delegation, feedback and taste take its place. The change measurably reorders who the top performers are — rank stability among engineers collapsed from 0.70 to 0.45 after AI, with the movement concentrated at staff and principal level. It reads more like a mindset switch than a tool skill, and once someone is classified an AI top performer they tend to stay there.

The person who cares about a decision and has authority to make it can now also execute it, so idea → design → spec → sprint queue → review compresses from days into minutes. “AI-native” in this reading is an org-design property, not a tooling one; product managers opening pull requests is the observable symptom rather than the definition.

Both shifts come with a stated line: the non-technical contributor owns user-facing change, the engineer keeps architecture. And on the widely-cited eight-level maturity ladder, level 8 is full agent orchestration — “I don’t know anybody who could credibly say they’re doing level 8 safely.” Being at level 2 or 3 is fine.

Also: Steve Yegge’s agentic-coding maturity ladder; Sophie Weston’s “broken comb”; incident.io on product engineers.

Assay’s authors are agents driven by briefs, not non-technical staff, so the “everyone a builder” half of this concept has no implementation here and none is planned. What does transfer is the boundary discipline: the rule that non-technical contributors own user-facing change while engineers keep architecture judgment is the same shape as Assay’s core-system rule, which decides on the surface touched rather than on who is asking. And the empowered-PM failure mode is the harness-blind defect class again — a specification authored outside the repository, unversioned, invisible to every gate.

Authoring briefs, arbitrating between streams, and holding the merge gate — not writing code. The caution worth keeping in view is that nobody credibly claims safe level-8 operation, and Assay runs a standing multi-desk fan-out, which is a level-7-or-8 posture held together by the merge gate and the guardrails rather than by demonstrated autonomy. Nothing in this evidence measures whether that holds. The honest statement is that Assay is further up the ladder than anyone in the evidence claims to be, on a smaller and more constrained surface.

• Whether the rank-stability finding (0.70 → 0.45) is replicated on a second population; it is currently the only quantitative claim behind “orchestration reorders performance”.
• Whether any organization publishes a developer-to-PM ratio experiment with an outcome attached, rather than the ratios currently traded as anecdote.
• Whether anyone describes level-7 or level-8 operation with the failure data attached, which is what would make the “nobody safely” claim testable.