After the Programmer Divide: One Measure for People Who Build Through AI Agents

Abstract

The verified delegation horizon, amplification gain, a dated map of role names, and 41 registered predictions scored each year until 2031. A bookkeeper asks an agent for a reconciliation tool; a developer asks one to change a billing system. Each must accept a day's work and show it holds. By October 2026 lawyers were setting tasks for the same agents too. In about 400,000 sessions, vendor-verified success on code-producing work was 34% for programming occupations and 29% for others, though occupation is not education. No career ladder says what someone building through agents can be trusted to deliver. I propose a measure in place of a title. The verified delegation horizon is the largest coherent piece of work, in reference hours of a qualified specialist, that a person regularly hands to agents and accepts after an independent check the agents did not see. Amplification gain, accepted reference hours per hour of attention, is a verified throughput, not a causal estimate of speedup. Both are scored per domain, Hour to Quarter; a desk test corrected one threshold. Three large employers tried token counts and dropped them in 2026: a ranking removed, a leaderboard gamed, a budget spent. Checked work is harder to game by spending, and each remaining route meets a stated defence. A dated map records rival names (builder, agent manager, principal and others) without choosing. Background neutrality is the hypothesis that at an equal horizon a programming education stops predicting checked quality. The credential crossover point is the moment, domain by domain, when products of people without an IT education begin to overtake those of programmers under the same checks, after which degrees and seniority labels lose weight. Forty-one dated predictions, priced on 9 October 2026 at the median of three blind model forecasts, are registered on OSF and scored yearly until 2031; a study never run counts as a failure, and overtaking in one's own domain stands at 0.35. Version 5 (11 October 2026) adds Appendix G: a non-programmer cohort of 119 people (35 already directing agents) and a Prolific rater study of 119 professionals on reference hours (registered on OSF, DOI 10.17605/OSF.IO/V92R8), with the panel size the rater rule needs. Version of record: SSRN 7589918. Predictions registered on OSF: https://doi.org/10.17605/OSF.IO/VW6P9. Code: https://github.com/vadimchernets/delegation-horizon (release v1.1.0; Zenodo concept DOI 10.5281/zenodo.23267444).