Harness engineering reframed: Workflow Definitions, Editor Platform, PaoBub for human-AI consensus

The author starts from a shift already under way in software development: humans used to write code with AI assisting, but increasingly AI implements, tests, and fixes while humans verify results. In that arrangement, the hard problem is no longer only whether the model can act, but whether the human and AI share enough understanding to agree on what should be built. OpenAI's and Anthropic's published work on agent harnesses shows that environment, constraints, and feedback matter — then asks whether that lens can be widened.

The proposed answer is deliberately minimal. Workflow Definitions reduces work to a few shared concepts (phase, deliverable, item, rule, acceptance), illustrated with both software development and manga production, and keeps references between items explicit so that changes upstream can be traced downstream. Editor Platform supplies the means to represent what text alone cannot — such as a storyboard with per-character timelines and locations — and aims to let humans and AI jointly build new editors when needed. PaoBub sits on top, tracking versions, approvals, and the impact of a changed judgment, so that prior sign-offs are not silently treated as still valid.

The author is candid that workflow engines, requirements management, and CSCW have long histories, and that the differences claimed here — including humans as first-class participants, prioritizing consensus information over execution steps, and making tools themselves modifiable — remain to be validated. Specific open questions include how small the necessary context can be, whether the same structure works across domains, what it costs to build custom tools, and how to assess the quality of consensus itself.