AI Agents Writing Code That Says 'Don't Use AI'
The qm project, a multiplayer agent harness for work, drew attention for an ironic detail in its contributing guidelines: human-written text only, no AI-generated code contributions, because 'coding agents write most underlying code now.' Commenters found this funny and revealing at the same time.
The pattern here is real: teams building on top of AI agents are discovering they need humans to anchor the intent layer, even when the execution layer is fully automated. The project uses Hermes as its underlying agent and is designed for teams running workflows across multiple AI workers in parallel. But the decision to require human-authored design documents shows that the people building these systems don't fully trust AI to reason about what should be built, only how to build it.
This connects directly to the broader 'prototype vs. product' debate. The judgment layer, the part where someone decides what the system should actually do, remains stubbornly human. Projects that pretend otherwise tend to drift into impressive demos that nobody can maintain.
So what?
If you're building an agent-based product or using agents internally, the governance layer matters as much as the capability layer. Being explicit about where humans must stay in the loop, and building that into your contribution or workflow process, is not a limitation. It's the thing that keeps the system coherent over time.