The Agentic Development Protocol: A Self-Falsifying Process Standard for Shipping Production Software with AI Agent Teams

Teams adopting AI coding agents report a consistent paradox: velocity rises while trust falls. In a randomized controlled trial, experienced developers took 19% longer to complete real tasks with early-2025 AI tools, while estimating after the fact that AI had made them 20% faster; in the same year, Stack Overflow's Developer Survey found that 66% of respondents cite AI output that is "almost right, but not quite" as a leading frustration. We argue the failure mode is methodological, not technical: teams run a single-typist methodology with faster typists, and no mechanism exists to verify what agent teams claim to have done. We present the Agentic Development Protocol (ADP), a production-derived experience report and process standard for running parallel AI agent sessions against one repository, extracted from a production system (1,768 commits, ~80 deploy rounds, 200+ closed tasks) rather than designed a priori. ADP's distinguishing property is that it is self-falsifying: every rule is earned through a numbered production miss, every claim carries an evidence grade with an explicit occurrence counter, promotion thresholds scale with blast radius, and the protocol publishes its own falsification record — 249 accepted claims and 43 falsified, counted by a re-runnable script against a pinned commit (§5.2), superseding the underived 529/22 pair asserted in earlier drafts — alongside its successes. We report an early within-researcher cross-stack replication attempt on a second, deliberately opposite install (deterministic embedded C++ vs. an LLM web stack), including a verification dimension we did not find addressed in prior frameworks: plan-phase semantic verification of task specifications. We state all limits explicitly: the sample is small, self-reported, and locally verified only. We therefore do not ask the reader to accept the internal numbers. ADP is released as an open protocol — markdown files, role prompts and optional hooks that install into an existing repository without replacing its orchestrator — so that the decisive evidence is generated by the teams who adopt it. The artifact is the protocol and the install ledger it accumulates; the paper's purpose is to get the hypothesis into other people's production environments, where it can be falsified. Artifact. The protocol, role prompts, hooks, ledger schemas, metric and claim-count scripts, and the full claim receipts (855 rows) are released at https://github.com/kikenandez/agentic-development-protocol — install reports, corroborating or disconfirming, are collected there. Version note. This is draft v0.5, a working preprint. §§4 and 6 retain outline markers where passages await source detail. Later versions will be published to this same Zenodo record.