Omega8 Reassesses AI Agents After BOA Documentation Errors

The team behind the BOA Drupal and Backdrop hosting stack says a failed AI-generated documentation attempt changed how it evaluates coding-agent output. In its 19 August 2026 blog post, Omega8 recounts using Claude Code and Dynamic Workflows to tackle documentation gaps accumulated across roughly 16 years of BOA development. The company says an initial run using Claude Opus 4.8 produced poorly structured material, and that a later Fable 5 pass exposed invented variables, incorrect interpretations, and descriptions of logic that did not exist.

Omega8 says its later workflow required documentation claims to point back to the code they described rather than treating generated prose as sufficient evidence. The resulting BOA documentation is divided into separate guides for hosted users, self-hosters, advanced operators, and developers. Anthropic separately describes Dynamic Workflows as a Claude Code feature that can break large tasks into parallel subagent work and coordinate the results. Omega8's assessment that Fable 5 performed substantially better than Opus 4.8 on its documentation work remains the company's own experience rather than a general comparison of model quality.

Omega8 says it has since expanded AI-assisted work into debugging, end-to-end test orchestration, and review of intrusion-detection rules used around hosted Drupal and Backdrop systems. The post also describes a shared library of session notes and a local documentation system created to preserve context across interdependent projects when individual Claude Code sessions could no longer track the environment reliably. Readers can examine the full progression, model commentary, and operational details in Artificial Stupidity, Revisited.